Protease data prediction method for catalytic reaction and application

Abstract

The application discloses a kind of proteinase data prediction methods and application for catalytic reaction, the method includes: obtaining protein sequence data, pocket mask and the molecular structure data of chemical reaction;Protein sequence data is input into first encoder to obtain protein feature data, and pocket feature data is extracted according to pocket mask;Molecular structure data is input into second encoder to obtain reaction feature data;Each feature data is mapped to hyperbolic space using learnable hyperbolic mapping parameter;Based on hyperbolic distance calculation comparative loss value, based on geometric relationship calculation implication loss value;Update encoder parameter and hyperbolic mapping parameter, obtain protein prediction model;Protein hyperbolic representation data and / or reaction hyperbolic representation data are output using the model, for enzyme catalytic function prediction, candidate enzyme and molecular structure data screening or enzyme sequence generation.The application realizes that only from sequence enzyme function accurate prediction and zero sample enzyme design.

Description

Technical Field

[0001] This invention relates to the fields of deep learning, protein engineering, and biocatalysis, and more specifically to technical solutions for using artificial intelligence technology to predict the properties and functions of proteases (especially industrial enzymes) and generate enzyme sequences. Background Technology

[0002] Proteases (enzymes) are widely distributed in organisms, catalyzing biochemical reactions and forming the core of biological metabolism. In industrial production, enzyme-catalyzed reactions have advantages such as high selectivity, high efficiency, and environmental friendliness, and are widely used in pharmaceuticals, chemicals, and food processing. However, a deep understanding and accurate prediction of enzyme properties remain major challenges in this field.

[0003] Traditional protein research mainly relies on experimental methods, such as X-ray crystallography and cryo-electron microscopy, to understand the function of proteins by resolving their three-dimensional structures. However, these methods face two key difficulties: (1) obtaining structural data is difficult and costly, and the three-dimensional structures of a large number of enzymes have not yet been resolved; (2) the correlation between structural information and enzyme catalytic activity is complex and difficult to predict directly.

[0004] In recent years, the emergence of large-scale protein models (such as the ESM series) has brought revolutionary progress to protein understanding. The ESM2 model (Evolution Scale Modeling), a pre-trained protein language model released by Meta AI, is trained using Masked Language Modeling (MLM) and can extract high-quality feature representations from protein sequences. This model performs exceptionally well on multiple protein understanding tasks, including secondary structure prediction and protein folding prediction.

[0005] Alongside large protein models, molecular coding techniques are being developed. ChemBERTa2 is a pre-trained model for SMILES (Simplified Molecular Input Line Entry System), which can represent chemical molecules as vectors, allowing chemical molecules to be compared and analyzed in vector space.

[0006] However, existing general-purpose large protein models have a fundamental limitation: they are trained based on general protein sequence data, and still have the following core problems in predicting enzyme properties and catalytic reactions:

[0007] Problem 1: General protein models have insufficient understanding of the catalytic properties of enzymes.

[0008] Existing general-purpose large protein models (such as the ESM series) are primarily optimized for general protein understanding tasks, such as sequence similarity prediction and secondary structure prediction. The training objectives of these models deviate significantly from the requirements for predicting enzyme catalytic properties. Therefore, directly using the representations of these models to predict characteristic indicators such as enzyme catalytic activity, optimal temperature, and optimal pH results in limited accuracy. While this can be improved through fine-tuning in downstream tasks, this approach requires a large amount of labeled data, which is often scarce for predicting specific enzyme properties.

[0009] Question 2: The use of structural information is limited.

[0010] In existing enzyme function prediction methods, researchers usually need to utilize the three-dimensional structural information of proteins to improve prediction accuracy. However, there is a serious problem of data scarcity in the acquisition of protein structure data: (1) the amount of structural data in PDB (protein database) is limited, and the structures of many industrial enzymes have not yet been resolved; (2) the experimental cost of obtaining protein structures is extremely high, and it is impossible to resolve the structures of all candidate enzymes; (3) even if structural data is available, the process of extracting and utilizing active site information is complex and prone to errors.

[0011] Question 3: Lack of systematic utilization of information on enzyme-catalyzed reactions.

[0012] Existing large-scale protein models primarily utilize evolutionary and structural information from protein sequences during training, but rarely consider information about the chemical reactions catalyzed by enzymes. However, the core function of enzymes is precisely to catalyze specific chemical reactions. Chemical reaction information (especially the substrate-enzyme-product tripartite information) can provide unique signals that guide the model to learn representations more specifically aligned with its catalytic function. Current methods fail to fully utilize this crucial information.

[0013] Question 4: Existing contrastive learning frameworks ignore the hierarchical structure of enzymes.

[0014] In recent years, existing protease-substrate contrastive learning frameworks in protein-molecule interaction research have primarily focused on enzyme discovery tasks, i.e., given a molecular reaction, retrieving enzymes from an enzyme library that may catalyze it. These methods share the following characteristics: Method principle: Using a contrastive learning framework, the enzyme sequence encoder and the molecular (substrate / product) encoder are mapped to a unified embedding space through projection layers in Euclidean space, and then the similarity is calculated for comparison.

[0015] Learning signal: Using known enzyme-reaction pairs as positive sample pairs and randomly sampled molecules as negative samples, training is performed by maximizing the similarity of positive pairs and minimizing the similarity of negative pairs.

[0016] The problem with this type of method is that it completely ignores the hierarchical structure of enzymes. Specifically: Multiple enzymes with the same catalytic pocket: In biology, many different enzymes may have very similar or even identical catalytic pocket configurations. For example, in enzyme classification, multiple enzymes within the same EC subclass often have similar active site geometry and catalytic mechanisms, and can catalyze the same type of chemical reaction or similar reactions.

[0017] Limitations of single similarity: Existing methods only calculate the global similarity between enzymes and molecules, completely ignoring the internal structural levels of enzymes. This leads to two enzymes with the same catalytic pocket being geographically distant in the embedding space, even if they actually catalyze the same reactions.

[0018] Information redundancy and repetitive learning: Because the pocket hierarchy is not distinguished, the model must learn almost the same mapping for each enzyme with similar functions, resulting in parameter redundancy and low learning efficiency.

[0019] Poor generalization performance: For new, unseen enzymes, the model has difficulty effectively utilizing learning information from other similar enzymes because the model does not explicitly represent and learn the universal characteristics of the active site.

[0020] Question 5: The multimodal alignment method is not optimized enough.

[0021] Existing multimodal alignment methods, such as protein-smiles contrastive learning, typically operate in Euclidean space, using a loss function based on dot product similarity. While concise, this approach overlooks the inherent hierarchical and inclusion relationships between proteins and molecules as different modalities. For example, a protein sequence and its catalytic pocket have a natural inclusion relationship, but simple dot product similarity cannot capture this relationship.

[0022] Question 6: It is impossible to design end-to-end enzymes solely from the sequence.

[0023] Because they rely on structural information, existing enzyme design methods typically require: (1) obtaining the three-dimensional structure of the target enzyme; (2) performing molecular docking simulations; and (3) performing structure-guided mutagenesis design. This process is computationally expensive, and its dependence on structural data severely limits its application. If a model could be established that allows understanding of enzyme properties based solely on sequence, it would revolutionize the de novo enzyme design process.

[0024] Therefore, how to improve the prediction accuracy of large protein models in enzyme catalysis applications is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0025] In view of the above problems, the present invention is proposed to provide a method and application for predicting protease data for catalytic reactions that overcomes or at least partially solves the above problems.

[0026] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a method for predicting protease data for catalytic reactions, characterized by comprising the following steps: Acquire training data, including: protein sequence data, pocket mask corresponding to the protein sequence data, and molecular structure data of chemical reactions, wherein the molecular structure data includes substrate molecular structure data and product molecular structure data; The protein sequence data is input into the first encoder to obtain protein feature data; pocket feature data is extracted from the protein feature data according to the pocket mask. The substrate molecular structure data and product molecular structure data are input into the second encoder to obtain reaction characteristic data; Using at least one learnable hyperbolic mapping parameter, the protein feature data, pocket feature data, and reaction feature data are respectively mapped to hyperbolic space to obtain protein hyperbolic representation data, pocket hyperbolic representation data, and reaction hyperbolic representation data; The contrast loss value is calculated based on the hyperbolic distance between the protein hyperbolic representation data and the reaction hyperbolic representation data, and the implied loss value is calculated based on the geometric relationship between the pocket hyperbolic representation data and the protein hyperbolic representation data. Based on the contrast loss value and the implied loss value, at least some parameters in the first encoder and the second encoder are updated, and the at least one learnable hyperbolic mapping parameter is updated to obtain the updated pedestal prediction model. Based on the protein sequence data to be processed and the pocket mask corresponding to the protein sequence data to be processed, and / or the molecular structure data of the chemical reaction to be processed, the updated protein pedestal prediction model outputs processed protein hyperbolic representation data and pocket hyperbolic representation data, and / or processed reaction hyperbolic representation data, for the prediction of protease or catalytic reactions.

[0027] Preferably, the step of extracting pocket feature data from the protein feature data based on the pocket mask includes: The features in the protein feature data that correspond to the positions with a value of 1 in the pocket mask are subjected to average pooling to obtain the pocket feature data; The pocket mask is a binary vector with the same length as the protein sequence data. A value of 1 indicates that the corresponding amino acid residue belongs to the active site.

[0028] Preferably, the step of mapping each feature data to a hyperbolic space using at least one learnable hyperbolic mapping parameter includes: By using a learnable projection matrix, each feature data is mapped to a latent space of a unified dimension, thereby obtaining feature data for each latent space. The latent spatial feature data are mapped to hyperbolic space through exponential mapping; The at least one learnable hyperbolic mapping parameter includes at least one of the following: The projection matrix parameters used to map feature data to the latent space; Curvature parameters used to control the degree of curvature in hyperbolic space; The scaling factor parameter is used to control the scaling degree when feature data is mapped to hyperbolic space.

[0029] Preferably, the total training loss function is constructed as follows:

[0030] in, To compare the loss values, To imply losses, It is a weight that implies loss.

[0031] Preferably, the comparison loss value is calculated as follows: Calculate the first hyperbolic distance between the hyperbolic representation data of the protein and the hyperbolic representation data of the positive sample reaction, wherein the positive sample is a chemical reaction that has a catalytic relationship with the protein sequence data; Calculate multiple second hyperbolic distances between protein hyperbolic representation data and multiple negative sample reaction hyperbolic representation data, wherein the negative samples are chemical reactions that do not have a catalytic relationship with the protein sequence data; Based on the first hyperbolic distance and the plurality of second hyperbolic distances, a cross-entropy loss is constructed to obtain the contrastive loss value.

[0032] Preferably, the implied loss value is calculated as follows: Calculate the implied angle between the pocket hyperbolic representation data and the protein hyperbolic representation data; Calculate the half-pore size of the hyperbolic representation data of the protein; Calculate the difference between the implied angle and the half-aperture, and take the maximum value between this difference and 0 to obtain the implied loss value.

[0033] Secondly, embodiments of the present invention provide a method for predicting enzyme catalytic function, comprising the following steps: Obtain the protein sequence data to be processed and the pocket mask corresponding to the protein sequence data to be processed; The protein sequence data to be processed and the pocket mask are input into the pedestal prediction model to obtain the processed hyperbolic representation data of the protein and the pocket hyperbolic representation data; wherein, the pedestal prediction model is obtained by the protease data prediction method for catalytic reactions described in the first aspect; The hyperbolic representation data and pocket hyperbolic representation data of the protein are input into a pre-trained prediction network, which outputs the enzyme catalytic function prediction results corresponding to the protein sequence data to be processed. The enzyme catalytic function prediction results include at least one of the following: the enzyme's EC number classification result, the optimal temperature value of the enzyme catalytic reaction, and the optimal pH value of the enzyme catalytic reaction.

[0034] Thirdly, embodiments of the present invention provide a method for predicting the kinetic parameters of an enzyme-catalyzed reaction, comprising the following steps: The protein sequence data of the target enzyme, the pocket mask corresponding to the protein sequence data of the target enzyme, and the molecular structure data of the target chemical reaction are obtained, wherein the molecular structure data includes substrate molecular structure data and product molecular structure data. The protein sequence data and pocket mask of the target enzyme are input into the pedestal prediction model to obtain enzyme hyperbolic representation data, which includes protein hyperbolic representation data and / or pocket hyperbolic representation data. The molecular structure data of the target chemical reaction is input into the pedestal prediction model to obtain hyperbolic representation data of the reaction; wherein, the pedestal prediction model is obtained by the protease data prediction method for catalytic reactions described in the first aspect; The hyperbolic representation data of the enzyme and the hyperbolic representation data of the reaction are fused and input into a pre-trained regression network to output the kinetic parameters of the target enzyme catalyzing the target chemical reaction.

[0035] Fourthly, embodiments of the present invention provide a method for bidirectional discovery of proteases and chemical reactions, comprising the following steps: Obtain query data, which is target enzyme data or target reaction data; wherein, the target enzyme data includes the protein sequence data of the target enzyme and its corresponding pocket mask, and the target reaction data includes the substrate molecular structure data and product molecular structure data of the target chemical reaction; The query data is input into the pedestal prediction model to obtain query hyperbolic representation data; wherein, the pedestal prediction model is obtained by the protease data prediction method for catalytic reactions described in the first aspect; when the query data is target enzyme data, the query hyperbolic representation data is enzyme hyperbolic representation data, including protein hyperbolic representation data and pocket hyperbolic representation data; when the query data is target reaction data, the query hyperbolic representation data is reaction hyperbolic representation data; Obtain a candidate set, which is an enzyme library or a reaction library; wherein, when the query data is target enzyme data, the candidate set is a reaction library, and each candidate reaction in the reaction library corresponds to pre-generated hyperbolic representation data of the reaction; when the query data is target reaction data, the candidate set is an enzyme library, and each candidate enzyme in the enzyme library corresponds to pre-generated hyperbolic representation data of the enzyme. Calculate the similarity between the query hyperbolic representation data and each candidate hyperbolic representation data in the candidate set; Candidate objects are sorted according to the similarity, and at least one candidate object is output based on the sorting results; wherein, when the query data is target enzyme data, the output candidate object is the candidate reaction that the target enzyme can catalyze; when the query data is target reaction data, the output candidate object is the candidate enzyme that catalyzes the target chemical reaction.

[0036] Fifthly, embodiments of the present invention provide a method for generating an enzyme protein sequence for a target chemical reaction, comprising the following steps: Obtain molecular structure data of the target chemical reaction, wherein the molecular structure data includes substrate molecular structure data and product molecular structure data; The molecular structure data of the target chemical reaction is input into the pedestal prediction model to obtain hyperbolic representation data of the target reaction; wherein, the pedestal prediction model is obtained by the protease data prediction method for catalytic reactions described in the first aspect; Obtain initial protein sequence data; By iteratively optimizing and adjusting the initial protein sequence data, the difference between the optimized protein sequence data and the target response hyperbolic representation data in the protein pedestal prediction model is reduced; When the preset convergence condition is met, the optimized protein sequence data is output as a candidate enzyme sequence generated for the target chemical reaction.

[0037] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: This invention provides a method and application for protease data prediction oriented towards catalytic reactions. By designing a peg prediction model specifically for enzymes, and maintaining the premise of using only protein sequences as input, a multimodal alignment of catalytic reaction information and catalytic pocket information is established to create an enzyme-specific peg prediction model. This enables the model to accurately predict enzyme properties and accurately understand catalytic reactions. In particular, explicit modeling of catalytic pockets allows the model to learn and represent shared catalytic features of different enzymes, thereby efficiently capturing the hierarchical nature of enzymes. Hyperbolic geometry is used for feature representation and alignment to fully capture the hierarchical relationships between different enzymes and efficiently utilize the shared features of multiple enzymes with the same or similar active site structures. A unique pocket masking and projection mechanism is designed to explicitly model the active site information of enzymes. Based on this peg prediction model, compared with existing general protein large models and existing protease-molecule comparative learning frameworks, accurate prediction of a series of downstream tasks is achieved, including single-modal enzyme property prediction and cross-modal reaction prediction.

[0038] As can be seen from the above technical solution, compared with the prior art, the present invention has significant advantages in at least the following aspects: Structure-independent: Completely sequence-based, eliminating dependence on the three-dimensional structure of proteins.

[0039] Hierarchical explicit modeling: Explicit modeling of catalytic pockets efficiently utilizes the shared features of multiple enzymes with the same function, overcoming the limitations of existing comparative learning methods.

[0040] Multi-faceted tasks: A single pre-trained model supports multiple downstream tasks, including property prediction, reaction prediction, enzyme discovery, and enzyme design.

[0041] Superior performance: Significant performance improvement (typically 20-50%) compared to general models such as ESM across various tasks, and 30-40% improvement in generalization performance compared to existing contrastive learning methods.

[0042] Theoretical innovation: The combined introduction of hyperbolic space, implication loss, and explicit pocket modeling is an academic innovation that brings new ideas to the application of deep learning in bioinformatics.

[0043] Practical value: It can directly support industrial applications and accelerate the discovery, modification, and engineering processes of enzymes.

[0044] Scalability: The model architecture is flexible and can be further extended to other protein engineering tasks. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0046] Figure 1 This is an overall architecture diagram of the base prediction model provided in this embodiment of the invention; Figure 2 This is a flowchart of the method for predicting enzyme catalytic function provided in the embodiments of the present invention; Figure 3 This is a flowchart of the method for predicting enzyme-catalyzed reaction kinetic parameters provided in the embodiments of the present invention; Figure 4 This is a flowchart of a method for discovering proteases involved in target chemical reactions provided in an embodiment of the present invention; Figure 5 This is a flowchart of the method for discovering reaction pairs in chemical reactions involving target proteases, provided in this embodiment of the invention. Figure 6 This is a flowchart of an enzyme protein sequence generation method for a target chemical reaction provided in an embodiment of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] This invention discloses a protease data prediction method for catalytic reactions. By constructing a pedestal prediction model and performing enzyme-reaction multimodal alignment in hyperbolic space, the model can understand the complex relationships between the enzyme, its catalytic pocket, and the catalyzed reaction. Based on this core design, the pedestal prediction model can be applied to multiple prediction tasks.

[0049] The overall workflow of the pedestal prediction model is as follows: Protein sequences obtain sequence-level contextual representations using an ESM2 encoder → Active site representations are extracted from the sequence representations using pocket masks → Protein and pocket representations are mapped to a unified latent space through a projection layer → The representations are projected to hyperbolic space through an exponential mapping. Molecules (substrate, product, and negative samples) are processed through a similar process. Finally, similarity is calculated in hyperbolic space and a loss function is constructed. Specifically, the steps include: Acquire training data, including: protein sequence data, pocket masks corresponding to protein sequence data, and molecular structure data of chemical reactions, including substrate molecular structure data and product molecular structure data; Protein sequence data is input into the first encoder to obtain protein feature data; pocket feature data is extracted from the protein feature data based on the pocket mask; Substrate molecular structure data and product molecular structure data are input into the second encoder to obtain reaction characteristic data; By using at least one learnable hyperbolic mapping parameter, protein feature data, pocket feature data, and reaction feature data are mapped to hyperbolic space respectively to obtain protein hyperbolic representation data, pocket hyperbolic representation data, and reaction hyperbolic representation data; The contrast loss value is calculated based on the hyperbolic distance between the protein hyperbolic representation data and the reaction hyperbolic representation data, and the implied loss value is calculated based on the geometric relationship between the pocket hyperbolic representation data and the protein hyperbolic representation data. Based on the comparison loss value and the implied loss value, at least some parameters in the first encoder and the second encoder are updated, and at least one learnable hyperbolic mapping parameter is updated to obtain the updated pedestal prediction model. Based on the protein sequence data to be processed and the pocket mask corresponding to the protein sequence data to be processed, and / or the molecular structure data of the chemical reaction to be processed, input to the updated protein pedestal prediction model, the model outputs processed protein hyperbolic representation data and pocket hyperbolic representation data, and / or processed reaction hyperbolic representation data, for the prediction of protease or catalytic reactions.

[0050] In one embodiment, two parallel encoders form a dual-encoder architecture: Protein encoder: Employs ESM2 (Evolution Scale Modeling) with a 1280-dimensional hidden layer representation. The input protein sequence is in a tokenized form, and the corresponding pocket mask uses a binary vector to mark which residues belong to active sites. The full-sequence protein encoding steps include: The protein sequence is first converted into a token sequence by the ESM2 tokenizer. Through forward propagation of the ESM2 encoder, a hidden state sequence of shape [batch_size, seq_length, 1280] is obtained. The hidden state of the CLS token (the first special marker of the sequence) is used as a global representation of the entire protein sequence, with a dimension of 1280. This representation captures global information about the entire protein sequence, including evolutionary history, expected secondary structures, etc., and forms the basis for unimodal tasks such as EC numbering and classification, and temperature / pH prediction.

[0051] Molecular encoder: ChemBERTa2 was used, optimized for SMILES representation, with a 768-dimensional hidden layer representation. Substrate and product SMILES were encoded and spliced ​​separately to obtain reaction-level molecular representations.

[0052] This dual encoder design allows for independent feature extraction of proteins and molecules, avoiding intermodal interference that may result from mixed encoding.

[0053] In one embodiment, explicit catalytic pocket modeling is introduced: Pocket masking mechanism: After protein encoding, the protein representation is masked using residue-level pocket masking to dynamically extract information around the active site.

[0054] Independent pocket projector: An independent projection layer is designed to map the pocket representation to the latent space, allowing the pocket representation to be optimized and adjusted independently.

[0055] This design enables the model to explicitly model active site information and efficiently utilize shared features among multiple enzymes in the same or similar catalytic pockets, thereby overcoming the problem of existing methods neglecting enzyme hierarchy.

[0056] In one embodiment, extracting pocket feature data from protein feature data based on a pocket mask includes: Average pooling is performed on the features in the protein feature data that correspond to the positions with a value of 1 in the pocket mask to obtain the pocket feature data; The pocket mask is a binary vector with the same length as the protein sequence data. A value of 1 indicates that the corresponding amino acid residue belongs to the active site.

[0057] The specific execution process of this embodiment is as follows: The catalytic pocket mask is a binary tensor of shape [batch_size, seq_length], where a value of 1 corresponds to an active site residue. The pocket representation is obtained by average pooling the hidden states at the masked locations.

[0058] in, This prevents small amounts from being divided by zero. This design allows for variable pocket size and automatic adaptation to the active site configuration of different enzymes. The pocket represents local information focused on the active site, which is particularly important for cross-modal tasks such as reaction kcat value prediction and enzyme-reaction matching, because catalytic reactions mainly occur at the active site rather than the entire protein.

[0059] In one embodiment, mapping feature data to a hyperbolic space using at least one learnable hyperbolic mapping parameter includes: By using a learnable projection matrix, each feature data is mapped to a latent space of a unified dimension, thereby obtaining feature data for each latent space. The latent spatial feature data are mapped to hyperbolic space through exponential mapping; At least one learnable hyperbolic mapping parameter includes at least one of the following: The projection matrix parameters used to map feature data to the latent space; Curvature parameters used to control the degree of curvature in hyperbolic space; The scaling factor parameter is used to control the scaling degree when feature data is mapped to hyperbolic space.

[0060] In this embodiment, the representations of proteins and pockets are mapped to a dimensionally uniform latent space (typically 512-dimensional) through independent linear projection layers:

[0061]

[0062] in and These are learnable projection matrices with dimensions [1280, 512] and [1280, 512], respectively. Similarly, the substrate and product codes for molecular SMILES are obtained using a ChemBERTa2 encoder (768 dimensions each), and after being concatenated, they are mapped to 512 dimensions through a projection layer.

[0063] The design of the projection layer has the following key functions: Feature Space Unification: Representations from different pre-trained models (1280-dimensional for ESM2, 768-dimensional for ChemBERTa2) are mapped to a unified 512-dimensional space. This allows features from different modalities to be meaningfully compared and aligned in a common feature space, especially for cross-modal tasks in Mode 2 (such as enzyme discovery given a reaction), where it is necessary to calculate the similarity between reaction and enzyme representations.

[0064] Dimensionality compression and feature learning: By mapping high-dimensional features to a relatively low dimension (512 dimensions), both computational complexity and the overhead of hyperbolic space mapping are reduced, while also facilitating the learning of more generalized and robust features. For single-modal prediction in application mode one (such as EC classification and temperature prediction), this compressed feature is already able to retain sufficient catalytic information.

[0065] Application-oriented feature learning: The weights of the projection layer are learned during training, enabling the model to automatically learn the feature representations best suited for multimodal contrastive learning objectives. These learned features naturally contain catalysis-related information, thus resulting in superior performance on downstream unimodal prediction tasks.

[0066] In this embodiment, to preserve and enhance the hierarchical structure of enzymes (different enzymes share similar catalytic pockets, etc.), all features are mapped to hyperbolic space. This is crucial for cross-modal prediction and de novo protein design, as the hierarchy of hyperbolic space can more efficiently represent the similarity between enzymes.

[0067] The mathematical basis of hyperbolic mappings is exponential mappings: In the Lorentz model, the exponential mapping originating from the origin maps vectors in Euclidean space. Mapped to points in hyperbolic space. The mapping formula is:

[0068] in It is the curvature of hyperbolic space. It is a scaling factor. It is the origin in the Lorentz model. In the implementation, a learnable scaling factor is introduced for numerical stability. (protein) and (molecular):

[0069]

[0070] Similarly, pocket and molecular features are treated.

[0071] The advantage of hyperbolic space lies in its inherent hierarchy compared to Euclidean space, with its distance function growing exponentially rather than nonlinearly. This means that in hyperbolic space, hierarchically related enzymes (such as enzymes within the same EC subclass) will automatically cluster together, while distant enzymes will be dispersed. This property directly benefits the "enzyme discovery given a reaction" task in Mode 2: when searching for enzymes similar to a given reaction, similarity calculations in hyperbolic space will prioritize returning enzymes with similar catalytic functions.

[0072] curvature The curvature was designed as a learnable parameter, with an initial value of 1.0. During training, the curvature was constrained within a reasonable range: (generally , Through this design, the model can adaptively select the hyperbolic curvature that best suits the current data. Higher curvature corresponds to a more curved space, which is better suited for representing highly hierarchical data. For enzyme data with a well-defined classification system (such as EC classification), the model automatically learns a higher curvature; for more uniformly distributed data, the curvature is adjusted to a lower value. This adaptability allows the model to optimally handle different data distributions.

[0073] In one embodiment, the HyperCLIP loss function is introduced, which contains two key components: Contrastive Loss: Uses cross-entropy loss to perform contrastive learning in hyperbolic space, making positive pairs more similar than negative pairs.

[0074] Entailment Loss: Utilizing the properties of hyperbolic geometry, this method forces pocket features to "entail" the entire protein feature in hyperbolic space, forming a natural hierarchical relationship.

[0075] In this embodiment, the comparison loss value is calculated as follows: Calculate the first hyperbolic distance between the hyperbolic representation data of proteins and the hyperbolic representation data of positive sample reactions, where positive samples are chemical reactions that have a catalytic relationship with the protein sequence data; Calculate multiple second hyperbolic distances between protein hyperbolic representation data and multiple negative sample reaction hyperbolic representation data, where negative samples are chemical reactions that have no catalytic relationship with the protein sequence data; Based on the first hyperbolic distance and multiple second hyperbolic distances, a cross-entropy loss is constructed to obtain the comparative loss value.

[0076] The comparison loss construction process in this embodiment is as follows: In the Lorentz model, two points The hyperbolic distance between them is defined as:

[0077] in It is the Lorentz inner product (a variant of the Minkowski inner product), defined as:

[0078] The key property of this distance is that, in hyperbolic space, distance growth is exponential rather than nonlinear, making the distance differences between distant points more apparent, which is suitable for representing hierarchical data. This property is particularly useful in the "enzyme discovery given a reaction" application mode two: due to the exponential growth of distance, the closest candidate enzyme is usually much smaller than other enzymes, resulting in higher accuracy for top-k searches.

[0079] For a batch of protein SMILES (substrate or product) The model calculates the distance between the protein and the positive sample. At the same time, calculate the protein and... SMILES of negative samples distance .

[0080] Based on distance, construct logarithmic probabilities:

[0081] in This is a temperature parameter. The contrast loss is cross-entropy loss, and the goal is to maximize the probability of the first position (facing the target).

[0082] in The similarity between corresponding proteins and molecules, This corresponds to the similarity between the pocket and the molecule. Note that two contrastive objectives are optimized simultaneously here, i.e., dual-contrast learning. The importance of dual-contrast learning lies in the fact that the global protein representation is used for single-modal prediction tasks (application mode 1), while the pocket representation is used for cross-modal tasks (application mode 2). By optimizing both objectives simultaneously, both representations are fully trained.

[0083] It should be noted that in each training step, negative samples are sampled from either the batch or a pre-computed pool to ensure the model continuously learns to distinguish difficult-to-distinguish negative pairs. Specifically, for each enzyme-reaction pair in each batch... Randomly sample from the global SMILES pool One (usually) Different molecules are used as negative samples. Sampling is dynamic, with resampling every epoch to prevent the model from overfitting to a fixed set of negative samples. As training progresses, the loss from easy negative samples (molecules that differ significantly from positive samples or known reactions) gradually decreases. The model automatically focuses on negative samples that are similar to positive samples (based on EC scales in the dataset to determine similarity), thus achieving the effect of sampling difficult negative samples. This further enhances the model's generalization ability. Substrates and products in multiple negative samples are also input in SMILES format.

[0084] In this embodiment, the implied loss value is calculated as follows: Calculate the implication angle between pocket hyperbolic representation data and protein hyperbolic representation data; Calculate the half-pore size of protein hyperbolic representation data; Calculate the difference between the implied angle and the half-aperture, and take the maximum value between this difference and 0 to obtain the implied loss value.

[0085] The implied loss construction process in this embodiment is as follows: By utilizing the properties of hyperbolic geometry, the pocket should be explicitly constrained to be "contained" or "encapsulated" by the protein. This design is particularly important for de novo enzyme design tasks. In hyperbolic space, a point The aperture reflects the range of points around a given point that it can encompass. Two points The relationship between them can be measured by their implication angle.

[0086] Definition of implied angle: Given two points (pocket) and protein The implied angle is defined as the geometric angle between them in hyperbolic space:

[0087] Definition of semi-aperture: protein dots The semi-aperture is defined as the geometric measure of the hyperbolic volume at that point: .

[0088] The core idea of ​​inclusion loss is that, biologically, the active site (pocket) is part of the whole protein and should be contained within it. This constraint is crucial for de novo enzyme design: when generating new protein sequences, the resulting sequences must satisfy this hierarchical relationship—that is, its pocket representation should be contained within the whole protein representation. Otherwise, the generated sequences may appear reasonable in sequence space but are biologically unreasonable (e.g., the relationship between the active site and the whole protein is disordered).

[0089] The mathematical definition of implication loss is: if the pocket is properly contained within the protein, the implication angle should be less than or equal to the protein's pore size. Implication loss is defined as:

[0090] in, It is a hyperparameter (usually 1.2) that controls how loosely the protein is contained in the pocket.

[0091] In one embodiment, the constructed total training loss function is:

[0092] in, To compare the loss values, To imply losses, This is the weight that implies the loss (typically 0.5). This combination ensures that the model learns both accurate similarity relationships, supporting enzyme-reaction matching prediction tasks, and the correct hierarchical structure, supporting de novo protein design tasks.

[0093] In one embodiment, to enable efficient fine-tuning of the model, a Low-Rank Adaptation (LoRA) mechanism is introduced: the weights of the pre-trained encoder are frozen, and only the projection layer, loss function parameters, and LoRA adapter are trained, which significantly improves the transferability and adaptability of the model.

[0094] The prediction method provided in this embodiment of the invention further includes the step of using the output hyperbolic representation data to achieve the following prediction tasks with important practical application value: First, it can accurately predict various important properties of enzymes, including their EC classification, optimal reaction temperature, and optimal pH. Second, given an orphan reaction (a reaction lacking a known catalytic enzyme), it can discover potential catalytic enzymes from existing enzyme libraries. Third, given a specific chemical reaction, it can de novo design a protein sequence capable of catalyzing the reaction, achieving zero-shot enzyme generation. Fourth, given a known catalytic enzyme, it can achieve rational optimization and modification of the enzyme by predicting the kinetic parameters (such as kcat value) after mutation.

[0095] These functions have significant application value in fields such as industrial biocatalysis, drug development, and environmental protection technologies. Specifically, this invention can significantly accelerate the discovery and design cycle of enzymes, reduce R&D costs, and provide strong technical support for solving many social problems (such as the development of biodegradable plastics and carbon capture). The following are implementation steps for different prediction tasks: In one embodiment, it further includes a step of predicting enzyme catalytic function by having the output hyperbolic representation data participate in the analysis.

[0096] In this embodiment, as Figure 2 As shown, the protein sequence data to be processed and the corresponding pocket mask are input into the updated pedestal prediction model to obtain the processed protein hyperbolic representation data and pocket hyperbolic representation data. The protein hyperbolic representation data participates in the prediction of enzyme catalytic function, including the enzyme's EC number classification. The execution steps include: The hyperbolic representation data and pocket hyperbolic representation data of proteins are input into a pre-trained classification network, which outputs the classification results of the enzyme EC number corresponding to the protein sequence data to be processed.

[0097] In this embodiment, as Figure 2As shown, the protein sequence data to be processed and the corresponding pocket mask are input into the updated pedestal prediction model to obtain the processed hyperbolic representation data and pocket hyperbolic representation data. The protein hyperbolic representation data and pocket hyperbolic representation data are used to predict enzyme catalytic function, including the prediction of enzyme catalytic reaction conditions. The execution steps include: The hyperbolic representation data and pocket hyperbolic representation data of proteins are input into a pre-trained regression network, which outputs the optimal temperature and / or optimal pH value of the enzyme-catalyzed reaction corresponding to the protein sequence data to be processed.

[0098] In this embodiment, as Figure 3 As shown, the protein sequence data to be processed and the corresponding pocket mask are input into the updated pedestal prediction model to obtain processed protein hyperbolic representation data and pocket hyperbolic representation data. Similarly, the molecular structure data of the chemical reaction to be processed are input into the updated pedestal prediction model to obtain processed reaction hyperbolic representation data. The protein hyperbolic representation data, pocket hyperbolic representation data, and reaction hyperbolic representation data are then used in the prediction of enzyme catalytic function, including the prediction of the enzyme's catalytic reaction rate constant. The execution steps include: The hyperbolic representation data of proteins, the hyperbolic representation data of pockets, and the hyperbolic representation data of reactions are fused together and input into a pre-trained regression network. The output is the catalytic rate constant value of the enzyme catalyzing the chemical reaction to be processed, corresponding to the protein sequence data to be processed.

[0099] In one embodiment, such as Figure 4 As shown, after obtaining the molecular structure data of the chemical reaction to be processed, inputting the molecular structure data of the chemical reaction to be processed into the updated pedestal prediction model, and obtaining the processed hyperbolic representation data of the reaction, the process also includes: a screening and prediction step of the output hyperbolic representation data participating in the enzyme catalyzing the target chemical reaction. The execution steps include: Obtain a candidate enzyme set, each candidate enzyme in the set corresponding to pre-generated hyperbolic representation data of the enzyme, including protein hyperbolic representation data and pocket hyperbolic representation data; Calculate the similarity between the hyperbolic representation data of the reaction and the hyperbolic representation data of each candidate enzyme; Candidate enzymes are ranked according to similarity, and at least one of the top-ranked candidate enzymes is output as a candidate enzyme to catalyze the chemical reaction to be treated.

[0100] In this embodiment, the hyperbolic representation data of enzymes in the candidate enzyme set are pre-generated in the following manner: The protein sequence data of the candidate enzyme and its corresponding pocket mask are input into the updated pedestal prediction model to obtain the processed protein hyperbolic representation data and pocket hyperbolic representation data, which are used as the enzyme hyperbolic representation data of the candidate enzyme.

[0101] In one embodiment, such as Figure 5 As shown, after obtaining the protein sequence data to be processed and the corresponding pocket mask, and inputting them into the updated pedestal prediction model to obtain the processed enzyme hyperbolic representation data (protein hyperbolic representation data and pocket hyperbolic representation data), the process also includes: a screening and prediction step for the molecular structure data of the output enzyme hyperbolic representation data participating in catalytic chemical reactions. The execution steps include: Obtain a candidate hyperbolic reaction set (substrate and product reaction pairs), where each candidate reaction pair corresponds to pre-generated hyperbolic reaction representation data; Calculate the similarity between the hyperbolic representation data of the enzyme and the hyperbolic representation data of each candidate reaction; The candidate reaction hyperbolic representation data are sorted according to similarity, and at least one candidate reaction hyperbolic representation data with the highest ranking is output as a candidate reaction pair for catalyzing the chemical reaction to be treated.

[0102] In this embodiment, the reaction hyperbolic representation data in the candidate reaction hyperbolic set is pre-generated in the following manner: The molecular structure data of the candidate chemical reaction is input into the updated pedestal prediction model to obtain the processed hyperbolic representation data of the reaction, which serves as the hyperbolic representation data of the molecular structure data of the candidate chemical reaction.

[0103] In one embodiment, such as Figure 6 As shown, the process involves acquiring molecular structure data of the target chemical reaction, inputting this data into an updated pedestal prediction model to obtain processed hyperbolic representation data of the reaction, and then further including a prediction step where the output hyperbolic representation data participates in the generation of enzyme sequences for the target chemical reaction. The execution steps include: Obtain initial protein sequence data; By iteratively optimizing and adjusting the initial protein sequence data, the difference between the optimized protein sequence data and the target response hyperbolic representation data in the updated pedestal prediction model is reduced. When the preset convergence conditions are met, the optimized protein sequence data is output as candidate enzyme sequences generated for the target chemical reaction.

[0104] In this embodiment, a step of quality assessment of the generated candidate enzyme sequence is also included: The candidate enzyme sequence is input into the protein structure prediction model to obtain the predicted three-dimensional structure and confidence score. Candidate pocket structures are extracted from the predicted 3D structure, and the similarity between the candidate pocket structures and the known active site structures is calculated. The overall quality score is calculated based on the confidence score and similarity score.

[0105] In this embodiment, a step of rationally optimizing the generated candidate enzyme sequence is also included: Mutate at least one amino acid site in the candidate enzyme sequence to generate sequence data of at least one mutant; The sequence data of each mutant and its corresponding pocket mask, as well as the molecular structure data of the target chemical reaction, are input into the updated pedestal prediction model to obtain the predicted kcat value of each mutant. Mutants with predicted kcat values ​​higher than the original kcat values ​​of the candidate enzyme sequence were selected as optimized candidate enzyme sequences.

[0106] Based on the same inventive concept, embodiments of the present invention also provide a method for predicting enzyme catalytic function, comprising the following steps: Obtain the protein sequence data to be processed and the pocket mask corresponding to the protein sequence data to be processed; The protein sequence data to be processed and the pocket mask are input into the pedestal prediction model to obtain the processed hyperbolic representation data of the protein and the pocket hyperbolic representation data; wherein, the pedestal prediction model is obtained by the aforementioned protease data prediction method for catalytic reactions. The hyperbolic representation data and pocket hyperbolic representation data of proteins are input into a pre-trained prediction network, which outputs the prediction results of enzyme catalytic function corresponding to the protein sequence data to be processed. The enzyme catalytic function prediction results include at least one of the following: the enzyme's EC number classification result, the optimal temperature value of the enzyme catalytic reaction, and the optimal pH value of the enzyme catalytic reaction.

[0107] Since this prediction method uses the pedestal prediction model from the aforementioned protease data prediction method for catalytic reactions, the implementation of this pedestal prediction model can be found in the implementation of the aforementioned method, and the repetition will not be repeated.

[0108] The specific steps for predicting enzyme catalytic function are explained step by step below: EC Number Prediction: The EC number is the standard classification code for enzymes, consisting of four numbers representing different enzymatic classification levels. (The text then abruptly shifts to a discussion of protein hyperbolic representation.) Input a classification network to predict the EC category of a sample. Compared to the general ESM2 model, this model performs better in EC prediction because it has fully learned catalytic information during training.

[0109] Predicting the optimal reaction temperature: The catalytic activity of enzymes is closely related to temperature, and different enzymes have different optimal operating temperatures. This is represented by the hyperbolic curve of the protein. The input is used to a regression network to predict the optimal temperature for the enzyme-catalyzed reaction. The prediction error of this model is reduced by 15-25% compared to ESM2 (measured by MAE).

[0110] Optimal pH prediction: Similarly, the optimal pH value is predicted using pocket features, which has a similar performance improvement compared to ESM2.

[0111] These predictions only require protein sequences as input, without any three-dimensional structural information.

[0112] Based on the same inventive concept, embodiments of the present invention also provide a method for predicting enzyme catalytic reaction kinetic parameters. Based on the multimodal alignment characteristics of the model in hyperbolic space, the following cross-modal prediction is achieved, including the following steps: Obtain the protein sequence data of the target enzyme, the pocket mask corresponding to the protein sequence data of the target enzyme, and the molecular structure data of the target chemical reaction. The molecular structure data includes the molecular structure data of the substrate and the molecular structure data of the product. The protein sequence data and pocket mask of the target enzyme are input into the pedestal prediction model to obtain enzyme hyperbolic representation data, which includes protein hyperbolic representation data and / or pocket hyperbolic representation data. The molecular structure data of the target chemical reaction is input into the pedestal prediction model to obtain hyperbolic representation data of the reaction; wherein, the pedestal prediction model is obtained through the aforementioned protease data prediction method for catalytic reactions; The hyperbolic representation data of enzymes and hyperbolic representation data of reactions are fused together and input into a pre-trained regression network to output the kinetic parameters of the target enzyme catalyzing the target chemical reaction.

[0113] Since this prediction method uses the pedestal prediction model from the aforementioned protease data prediction method for catalytic reactions, the implementation of this pedestal prediction model can be found in the implementation of the aforementioned method, and the repetition will not be repeated.

[0114] The following explains the steps for predicting specific enzyme-catalyzed reaction kinetic parameters: Reaction kinetic parameter prediction (kcat value): Given a specific enzyme and reaction pair, predict the rate constant kcat that the enzyme catalyzes for the reaction. Using the pocket representation of the enzyme and the molecular representation of the reaction, their similarity and relationship in hyperbolic space are calculated, and the kcat value is predicted using a fine-tuned regression head. This is of great significance for the rational optimization design of enzymes.

[0115] Based on the same inventive concept, embodiments of the present invention also provide a method for bidirectional discovery of proteases and chemical reactions, comprising the following steps: Retrieve query data, which can be either target enzyme data or target reaction data; target enzyme data includes the protein sequence data of the target enzyme and its corresponding pocket mask, while target reaction data includes the substrate molecular structure data and product molecular structure data of the target chemical reaction. The query data is input into the pedestal prediction model to obtain the query hyperbolic representation data; wherein, the pedestal prediction model is obtained through the aforementioned protease data prediction method for catalytic reactions; when the query data is target enzyme data, the query hyperbolic representation data is enzyme hyperbolic representation data, including protein hyperbolic representation data and pocket hyperbolic representation data; when the query data is target reaction data, the query hyperbolic representation data is reaction hyperbolic representation data; Obtain a candidate set, which can be an enzyme library or a reaction library. When the query data is target enzyme data, the candidate set is a reaction library, and each candidate reaction in the reaction library corresponds to pre-generated hyperbolic representation data of the reaction. When the query data is target reaction data, the candidate set is an enzyme library, and each candidate enzyme in the enzyme library corresponds to pre-generated hyperbolic representation data of the enzyme. Calculate the similarity between the query hyperbolic representation data and each candidate hyperbolic representation data in the candidate set; Candidate objects are sorted according to similarity, and at least one candidate object is output based on the sorting results; wherein, when the query data is target enzyme data, the output candidate objects are candidate reactions that the target enzyme can catalyze; when the query data is target reaction data, the output candidate objects are candidate enzymes that catalyze the target chemical reaction.

[0116] Since this discovery method uses the pedestal prediction model from the aforementioned protease data prediction method for catalytic reactions, the implementation of this pedestal prediction model can be found in the implementation of the aforementioned method, and the repetitions will not be repeated.

[0117] Based on the model's multimodal alignment characteristics in hyperbolic space, cross-modal prediction is achieved. The following describes the specific steps for bidirectional discovery of proteases and chemical reactions: Enzyme discovery: Given a chemical reaction (specifying substrates and products SMILES), encode the representation of that reaction in hyperbolic space. Then, similarity calculations are performed with all enzyme representations in the existing enzyme library to identify the enzymes that are closest in hyperbolic space as candidates. These candidate enzymes are most likely to catalyze the reaction. The retrieval accuracy (recall@10) reaches 70-80%, a significant improvement compared to the 40-50% of general models.

[0118] Given an enzyme, discover the reactions it can catalyze (reaction discovery): Conversely, given a specific enzyme sequence, its representation in hyperbolic space is encoded, and then similarity calculations are performed with all reaction representations in a known enzyme-reaction database to identify the reactions most likely to be catalyzed by that enzyme. This is significant for discovering novel enzyme functions. Performance is also improved by 30-40% compared to existing methods.

[0119] Based on the same inventive concept, embodiments of the present invention also provide a method for generating enzyme protein sequences for a target chemical reaction. The method for generating enzyme protein sequences for a target chemical reaction includes the following steps: Obtain molecular structure data for the target chemical reaction, including substrate molecular structure data and product molecular structure data; The molecular structure data of the target chemical reaction is input into the pedestal prediction model to obtain the hyperbolic representation data of the target reaction; wherein, the pedestal prediction model is obtained through the aforementioned protease data prediction method for catalytic reactions; Obtain initial protein sequence data; By iteratively optimizing and adjusting the initial protein sequence data, the difference between the optimized protein sequence data and the target response hyperbolic representation data in the protein pedestal prediction model is reduced. When the preset convergence conditions are met, the optimized protein sequence data is output as candidate enzyme sequences generated for the target chemical reaction.

[0120] Since this generation method uses the pedestal prediction model from the aforementioned protease data prediction method for catalytic reactions, the implementation of this pedestal prediction model can be found in the implementation of the aforementioned method, and the repetitions will not be repeated.

[0121] The following describes the steps for generating enzyme protein sequences for a target chemical reaction: Based on the enzyme-reaction pairing information learned by the model in the latent space, and the alignment relationship between the hyperbolic representation of proteins and the hyperbolic representation of reactions, zero-sample de novo enzyme design can be achieved. Protein sequence generation for a given reaction: Given a target chemical reaction, first encode the representation of the reaction in hyperbolic space. Then, using a gradient-guided sequence optimization method, starting from an initial protein sequence, the distance between the generated sequence's protein representation and the target response representation in hyperbolic space is minimized through backpropagation. Simultaneously, a sequence regularization term is introduced to ensure the generated sequence has biological plausibility (e.g., avoiding extreme amino acid compositions). After multiple iterations of optimization, a protein sequence highly aligned with the target response is obtained.

[0122] This design method is based entirely on the alignment of sequences and latent space, requiring no three-dimensional structural information or molecular docking calculations, which greatly reduces computational costs (from the traditional weeks or months to hours).

[0123] The generated protein sequence can be evaluated across multiple dimensions, including similarity to known catalytic enzymes of the reaction, foldability assessment via ESMFold (pLDDT score), and similarity to known active site structures. Sequences with high overall scores are more likely to be truly active enzymes.

[0124] The inventive concept upon which the above embodiments of the present invention are based involves designing a large model specifically for enzymes. While maintaining the use of only protein sequences as input, the model learns multimodal catalytic reaction information and catalytic pocket information, enabling accurate prediction of enzyme properties and precise understanding of catalytic reactions. Specifically, by explicitly modeling catalytic pockets and aligning them in hyperbolic space, the model can capture hierarchical relationships between different enzymes and efficiently utilize the shared features of multiple enzymes with the same or similar active site structures. Compared to existing general protein models and existing protease-molecule contrastive learning frameworks, the model of the present invention achieves significant performance improvements in multiple enzyme-related prediction tasks and effectively solves the problems existing in the prior art. The details are as follows: Question 1: General protein models such as ESM2 did not take into account the specificity of enzymes in their design, and have limited understanding of enzyme catalytic properties (such as optimal temperature, optimal pH, and EC classification).

[0125] This invention integrates a large amount of enzyme-reaction pair data during training, particularly utilizing structured information from the Rhea enzyme reaction database, enabling the model's feature representation to naturally incorporate catalytic information. During the inference phase, only protein sequences are needed for prediction.

[0126] The advantages of this invention are reflected in the following aspects: In EC classification and prediction tasks, the accuracy is improved by 20-30% compared to ESM2.

[0127] In regression predictions for optimal temperature and optimal pH, the mean absolute error (MAE) was reduced by 15-25%, respectively.

[0128] Only sequential input is required, completely avoiding dependence on three-dimensional structures.

[0129] Problem 2: Existing protease-molecular contrastive learning methods focus on enzyme discovery tasks, calculating global similarity in Euclidean space, completely ignoring the hierarchical structure of enzymes. In particular, multiple different enzymes with the same or similar catalytic pockets may be mapped to completely different embedding space locations, causing the model to fail to effectively utilize the shared features between these enzymes.

[0130] The solution of this invention includes: explicitly modeling catalytic pockets and designing independent pocket projection layers, allowing pocket features to be learned and optimized individually. Multimodal alignment is performed using hyperbolic space; the natural hierarchy of hyperbolic geometry makes multiple enzymes with the same catalytic function closer together in the embedding space. An inclusion loss is introduced, explicitly constraining whether the pocket representation should be "contained" or "encapsulated" by the protein representation, forming a natural hierarchical relationship.

[0131] The advantages of this invention are reflected in the following aspects: When clustering enzymes within the same EC subclass using pocket features, the intra-class distance is reduced by 13-20% compared to global similarity methods.

[0132] For novel enzymes that have not been seen before, the model can better utilize learning information from other known enzymes (especially those with similar pockets).

[0133] The model parameters are more efficient because there is no need to learn redundant mappings separately for enzymes with similar functions.

[0134] Question 3: Most industrial enzymes lack resolved three-dimensional structures, leading to a data shortage when using structure-based prediction methods.

[0135] The solution of this invention includes: explicit modeling of the catalytic pocket, enabling the model to learn key features of the active site at the sequence level without the need for external structural prediction. Through hyperbolic entailment loss, the inclusion relationship between the pocket representation and the complete protein representation is explicitly constrained, resulting in more accurate pocket features. A pocket masking mechanism is designed to directly extract pocket information from the protein sequence and known active site residue indices, without the need for 3D coordinates.

[0136] The advantages of this invention are reflected in the following aspects: It completely eliminates the dependence on the three-dimensional structure of proteins.

[0137] Even for enzymes where structural data is lacking, the model can still provide reliable predictions.

[0138] Pocket-related prediction tasks (such as response-specific identification) show significantly improved performance compared to using only whole-sequence features.

[0139] Question 4: There is a huge discrepancy between the training objectives of existing large protein models (such as MLM) and the actual function of enzymes (catalyzing specific reactions).

[0140] The solution of this invention includes: introducing explicit reaction information into model training: each protein is compared and learned in contrast with the substrates and products of all the reactions it catalyzes. A dual-contrast learning strategy is designed to learn both substrates and products separately, enabling the model to understand both the start and end points of the reaction. Sampling of negative samples ensures that the model can distinguish between real reactions and randomly paired molecules.

[0141] The advantages of this invention are reflected in the following aspects: The semantics learned by the model naturally possess catalytic functions and can be directly used for reaction prediction.

[0142] On the task of predicting enzyme reactions (enzyme function mining), the retrieval accuracy (recall@10) reaches 70-80%, which is a significant improvement compared to the 40-50% of the general model.

[0143] Similar performance improvements are achieved on tasks involving predicting enzymes for a given reaction (enzyme discovery of the reaction).

[0144] Question 5: Simple dot product similarity cannot capture the essential relationships between different modalities, especially inclusion relationships.

[0145] The solution of this invention includes: designing the HyperCLIP loss to compute similarity in hyperbolic space, using Lorentz distance instead of Euclidean distance. An implication loss is introduced, explicitly constraining that pockets should be contained within proteins, forming natural hierarchical relationships. The implication expression is achieved through the concepts of angle and aperture in hyperbolic geometry, allowing this relationship to be precisely quantified and optimized.

[0146] The advantages of this invention are reflected in the following aspects: The model learns more stable and robust representations. The model using implied loss performs better on noisy data compared to variants that only use contrastive loss.

[0147] The ternary relationship between protein-pocket-molecule is explicitly constrained and optimized, making the learned structure of the model more reasonable.

[0148] Question 6: Existing enzyme design methods rely heavily on structural information and complex calculations, making it difficult to achieve zero-sample enzyme design.

[0149] The solution of this invention includes seamlessly integrating the pedestal model with a protein generation model based on the latent space characteristics and reaction pairing information of the pedestal model. During the generation process, the representation of the generated protein sequence in the model's latent space is constrained to be similar to the representation of the target reaction, thereby generating the catalytic enzyme for a given reaction. This method is entirely sequence-based and requires no intermediate structure prediction steps.

[0150] The advantages of this invention are reflected in the following aspects: For the first time, zero-sample enzyme design for a specific chemical reaction has been achieved.

[0151] Compared to traditional methods (which take weeks to months), the method of this invention can generate candidate sequences in hours.

[0152] The generated sequences are more plausible and more likely to be foldable, expressive, and active.

[0153] To fully illustrate the execution process of the method of the present invention, the following is a description through specific examples: Case background and problem description: In the pharmaceutical and fine chemical industries, there are numerous orphan reactions—chemical reactions known but lacking known catalytic enzymes. Solving these reactions is crucial for the industry, but traditional methods face several challenges: natural screening is time-consuming and labor-intensive, chemical catalysis can cause pollution, and rational design requires structural information and incurs high computational costs.

[0154] This case demonstrates how to use the protease model of this invention to completely solve an orphan reaction problem, covering the three steps of enzyme discovery, enzyme design, and enzyme optimization.

[0155] The target industrial reaction is assumed to be the reduction of an aromatic ketone (phenylbenzyl ketone, substrate SMILES: `c1ccc(cc1)C(=O)c2ccccc2`) to the corresponding alcohol (product SMILES: `c1ccc(cc1)C(O)c2ccccc2`). This is a typical reduction reaction. A pre-trained protease locator model containing 260,000 enzyme-reaction pairs already exists; the task now is to address this orphan reaction.

[0156] Step 1: Discovering potential catalytic enzymes from the enzyme library. The specific steps are as follows: Objective: To rapidly identify candidate enzymes that may catalyze this reduction reaction from a known enzyme library (10,000 published enzymes).

[0157] Implementation process: (1) Encoding and representation of target response: The substrate and product SMILES were encoded using a ChemBERTa2 word segmenter and encoder to obtain their hidden state representations. The two representations were concatenated to form a molecular vector representing the reaction order, which was then mapped to a 512-dimensional latent space through a projection layer. Finally, it was projected onto a hyperbolic space using an exponential mapping to obtain the hyperbolic space representation of the target reaction. .

[0158] (2) Precoding of the known enzyme library: 10,000 enzymes were encoded offline. The sequence of each enzyme was processed using an ESM2 encoder to extract the full-sequence CLS token representation. This token was then mapped to 512 dimensions using a protein projection layer, and finally projected onto hyperbolic space using an exponential mapping. .

[0159] (3) Similarity search and ranking: For each enzyme in the library, calculate its Lorentz distance to the target reaction:

[0160] The smaller the distance, the higher the similarity. Sort all enzymes by distance from smallest to largest, and select the 20 enzymes with the smallest distance as candidates.

[0161] Enzyme library retrieval is performed using the following pseudocode: Algorithm: Enzyme Library Retrieval Input: Target reactions SMILES, enzyme library E, pre-trained model M Output: List of candidate enzymes 1. r_target = encoding response(target_SMILES, M) 2. Similarity = [] 3. Regarding enzyme e in E: e_rep = encoding enzyme(e,M) dist = calculate distance(r_target, e_rep) Similarity. Add (dist, ID of e) 4. Ranking Similarity (Ascending Order) 5. Return the top 20 enzymes with the lowest similarity. Expected Results: The model is expected to rapidly (within seconds) retrieve enzymes associated with reduction reactions from 10,000 enzymes. The top-ranked enzymes are anticipated to be known keto reductases (e.g., redox variants of certain alcohol dehydrogenases). This process is more than 100 times more efficient than blind screening (which requires large-scale, high-throughput experiments).

[0162] Step 2: Optimize or generate candidate enzymes from scratch, the specific steps are as follows: If the candidate enzymes retrieved above do not meet the criteria (e.g., low similarity, or known difficulty in expressing candidate enzymes), then the de novo enzyme design phase begins.

[0163] Objective: To generate a new protein sequence that can catalyze this reduction reaction.

[0164] Implementation process: (1) Hyperbolic representation of the target response: Similar to methods used in enzyme mining, a hyperbolic representation of the target reaction is obtained. .

[0165] (2) Initial protein sequence generation: Choose a protein sequence of reasonable length as initialization (this can be randomly generated or a template drawn from a known reductase). Perform one-hot encoding on this sequence to form an optimizable representation. ,in It is the sequence length. This is the size of the amino acid word list (usually 33).

[0166] (3) Gradient-guided sequence optimization: The optimization objective is defined as: minimizing the distance between the hyperbolic representation of the generated protein sequence and the target reaction representation, while ensuring that the generated sequence has reasonable biological characteristics.

[0167] The optimization process is executed via the following pseudocode: Algorithm: Gradient-guided protein sequence optimization Input: Initial sequence x_0, target response representation r_target, number of iterations N, learning rate lr Output: Optimized protein sequence 1. x = x_0, x.requires_grad = True 2. Optimizer = Adam(x, lr=lr) 3. For iter=1 to N: Protein representation = encoded protein(x) Protein hyperbolic = Mapping to hyperbolic (protein representation) Distance loss = Calculated distance(protein hyperbolic, r_target) Sequence regularization = regularization term(x) / / Encourages reasonable amino acid distribution Total loss = Distance loss + λ Sequence regularization Total loss. Backpropagation() optimizer.step() 4. Final sequence = obtaining the most likely amino acid from x. 5. Return the final sequence The purpose of the sequence regularization term is to prevent the optimization from getting stuck in an unreasonable sequence space, such as ensuring that all positions have valid amino acids (through entropy regularization).

[0168] (4) Quality assessment of generated sequences: Each generated sequence is evaluated across multiple dimensions: Catalytic relevance: Calculate its similarity score to the target reaction (directly from the model).

[0169] Foldability: Use ESMFold to quickly predict whether a sequence can be folded into a reasonable three-dimensional structure and calculate the pLDDT confidence score.

[0170] Pocket plausibility: Possible active sites were extracted from the predicted 3D structure and their similarity to known reductase pockets was assessed.

[0171] By combining these three dimensions, an overall quality score is calculated, and the sequence with the highest score is selected as a candidate.

[0172] Expected results: The model is expected to generate multiple candidate sequences within hours. The generated sequences should have: (1) a feature representation that is highly correlated with the target reaction, (2) a reasonably foldable secondary structure prediction, and (3) contain key catalytic residues similar to those of known reductases.

[0173] Step 3: Rational optimization of candidate enzymes. The specific steps are as follows: After obtaining candidate enzymes (through mining or generation), further rational sequence optimization is carried out to improve their catalytic efficiency (measured by kcat).

[0174] Objective: To identify advantageous single-point mutations through site mutation scanning, and further optimize the enzyme's catalytic rate constant (kcat).

[0175] Implementation process: (1) Construction of the kcat prediction model: Based on a pre-trained protease locus model, an auxiliary kcat regression model is constructed. This model takes hyperbolic representations of the protein and its pocket as input and predicts kcat values ​​using a small neural network regression head.

[0176] This model is trained in a supervised manner on a training set containing known enzyme-reaction kcat values.

[0177] (2) Site mutation scanning: For each site of the candidate enzyme (typically focusing on the active site region or the area surrounding the active site), try 20 standard amino acid substitutions one by one. For each mutant... Calculate the predicted kcat value. The advantage of the mutation is defined as the relative improvement:

[0178] in, This indicates the wild type (original candidate enzyme).

[0179] Site mutation scanning is performed using the following pseudocode: Algorithm: Site mutation scanning Input: Candidate enzyme sequence, kcat prediction model, active site range [start, end] Output: A list of mutations sorted by degree of improvement. 1. wt_kcat = predicted kcat (candidate enzyme) 2. Mutation list = [] 3. For pos=start to end: For the new amino acid aa in 20 amino acids: Mutant sequence = Replacement (candidate enzyme, pos, new aa) mut_kcat = predict kcat(mutated sequence) Improvement degree = (mut_kcat - wt_kcat) / wt_kcat Mutation list. Add(location, original aa → new aa, improvement level) 4. Sort the mutation list (in descending order of improvement) 5. Return to the top 10 most promising mutations (3) Prediction and verification of multi-point mutations: Based on the results of single-point scans, select the 3-5 single-point mutations with the best predictive performance and combine them to predict multi-point mutations. Verify whether these combinations have an additive effect (whether the combination of multiple beneficial mutations is more effective).

[0180] Expected results: The scan should identify 10-20 promising mutation sites, with the top 5 mutations potentially improving kcat by 10-30%. Ultimately, the single- or multi-site mutation scheme with the best overall performance will be selected for subsequent experimental validation.

[0181] This case demonstrates that the protease locus model described in this invention can completely solve an industrial orphan reaction problem: 1. Enzyme mining steps (first step): Compared with traditional high-throughput screening, the model can narrow down the candidate range from 10,000 to 20, saving a lot of screening costs and time.

[0182] 2. Enzyme design steps (second step): If none of the candidates in the library meet the conditions, the model can generate new candidate sequences from scratch, achieving "zero-sample" design, which is something that traditional methods cannot do.

[0183] 3. Enzyme optimization step (third step): Rational mutation optimization is performed on the most promising candidates to further improve their catalytic efficiency, which is more than 100 times more efficient than the random mutation method.

[0184] The total time taken for the entire workflow (from reaction to final candidate) is approximately several hours to several days (mainly depending on the number of candidates generated and the evaluation calculations), significantly accelerating the enzyme engineering cycle compared to the weeks to months of traditional methods. This fully demonstrates the practical value of this invention in industrial applications.

[0185] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0186] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting protease data for catalytic reactions, characterized in that, Includes the following steps: Acquire training data, including: protein sequence data, pocket mask corresponding to the protein sequence data, and molecular structure data of chemical reactions, wherein the molecular structure data includes substrate molecular structure data and product molecular structure data; The protein sequence data is input into the first encoder to obtain protein feature data; pocket feature data is extracted from the protein feature data according to the pocket mask. The substrate molecular structure data and product molecular structure data are input into the second encoder to obtain reaction characteristic data; Using at least one learnable hyperbolic mapping parameter, the protein feature data, pocket feature data, and reaction feature data are respectively mapped to hyperbolic space to obtain protein hyperbolic representation data, pocket hyperbolic representation data, and reaction hyperbolic representation data; The contrast loss value is calculated based on the hyperbolic distance between the protein hyperbolic representation data and the reaction hyperbolic representation data, and the implied loss value is calculated based on the geometric relationship between the pocket hyperbolic representation data and the protein hyperbolic representation data. Based on the contrast loss value and the implied loss value, at least some parameters in the first encoder and the second encoder are updated, and the at least one learnable hyperbolic mapping parameter is updated to obtain the updated pedestal prediction model. Based on the protein sequence data to be processed and the pocket mask corresponding to the protein sequence data to be processed, and / or the molecular structure data of the chemical reaction to be processed, the updated protein pedestal prediction model outputs processed protein hyperbolic representation data and pocket hyperbolic representation data, and / or processed reaction hyperbolic representation data, for the prediction of protease or catalytic reactions.

2. The protease data prediction method for catalytic reactions as described in claim 1, characterized in that, The step of extracting pocket feature data from the protein feature data based on the pocket mask includes: The features in the protein feature data that correspond to the positions with a value of 1 in the pocket mask are subjected to average pooling to obtain the pocket feature data; The pocket mask is a binary vector with the same length as the protein sequence data. A value of 1 indicates that the corresponding amino acid residue belongs to the active site.

3. The protease data prediction method for catalytic reactions as described in claim 1, characterized in that, The step of mapping each feature data to a hyperbolic space using at least one learnable hyperbolic mapping parameter includes: By using a learnable projection matrix, each feature data is mapped to a latent space of a unified dimension, thereby obtaining feature data for each latent space. The latent spatial feature data are mapped to hyperbolic space through exponential mapping; The at least one learnable hyperbolic mapping parameter includes at least one of the following: The projection matrix parameters used to map feature data to the latent space; Curvature parameters used to control the degree of curvature in hyperbolic space; The scaling factor parameter is used to control the scaling degree when feature data is mapped to hyperbolic space; In the Lorentz model, the exponential mapping originating from the origin maps vectors in Euclidean space. The mapping formula for points in hyperbolic space is: in, It is the curvature parameter of hyperbolic space. It is a scaling factor. It is the origin in the Lorentz model; Scaling factor parameters include protein scaling factor and molecular scaling factor Among them, for protein scaling factor : ； curvature parameters The constraint interval is .

4. The protease data prediction method for catalytic reactions as described in claim 1, characterized in that, The total training loss function is constructed as follows: in, To compare the loss values, To imply losses, It is a weight that implies loss.

5. The protease data prediction method for catalytic reactions as described in claim 1, characterized in that, The method for calculating the comparative loss value is as follows: Calculate the first hyperbolic distance between the hyperbolic representation data of the protein and the hyperbolic representation data of the positive sample reaction, wherein the positive sample is a chemical reaction that has a catalytic relationship with the protein sequence data; Calculate multiple second hyperbolic distances between protein hyperbolic representation data and multiple negative sample reaction hyperbolic representation data, wherein the negative samples are chemical reactions that do not have a catalytic relationship with the protein sequence data; Based on the first hyperbolic distance and the plurality of second hyperbolic distances, a cross-entropy loss is constructed to obtain the contrastive loss value.

6. The method for predicting protease data for catalytic reactions as described in claim 1, characterized in that, The implied loss value is calculated as follows: Calculate the implied angle between the pocket hyperbolic representation data and the protein hyperbolic representation data; Calculate the half-pore size of the hyperbolic representation data of the protein; Calculate the difference between the implied angle and the half-aperture, and take the maximum value between this difference and 0 to obtain the implied loss value.

7. A method for predicting enzyme catalytic function, characterized in that, Includes the following steps: Obtain the protein sequence data to be processed and the pocket mask corresponding to the protein sequence data to be processed; The protein sequence data to be processed and the pocket mask are input into the pedestal prediction model to obtain the processed hyperbolic representation data of the protein and the pocket hyperbolic representation data; wherein, the pedestal prediction model is obtained by the protease data prediction method for catalytic reactions according to any one of claims 1 to 6; The hyperbolic representation data and pocket hyperbolic representation data of the protein are input into a pre-trained prediction network, which outputs the enzyme catalytic function prediction results corresponding to the protein sequence data to be processed. The enzyme catalytic function prediction results include at least one of the following: the enzyme's EC number classification result, the optimal temperature value of the enzyme catalytic reaction, and the optimal pH value of the enzyme catalytic reaction.

8. A method for predicting kinetic parameters of enzyme-catalyzed reactions, characterized in that, Includes the following steps: The protein sequence data of the target enzyme, the pocket mask corresponding to the protein sequence data of the target enzyme, and the molecular structure data of the target chemical reaction are obtained, wherein the molecular structure data includes substrate molecular structure data and product molecular structure data. The protein sequence data and pocket mask of the target enzyme are input into the pedestal prediction model to obtain enzyme hyperbolic representation data, which includes protein hyperbolic representation data and pocket hyperbolic representation data. The molecular structure data of the target chemical reaction is input into the pedestal prediction model to obtain hyperbolic representation data of the reaction; wherein, the pedestal prediction model is obtained by the protease data prediction method for catalytic reactions according to any one of claims 1 to 6; The hyperbolic representation data of the enzyme and the hyperbolic representation data of the reaction are fused and input into a pre-trained regression network to output the kinetic parameters of the target enzyme catalyzing the target chemical reaction.

9. A bidirectional discovery method for protease and chemical reactions, characterized in that, Includes the following steps: Obtain query data, which is target enzyme data or target reaction data; wherein, the target enzyme data includes the protein sequence data of the target enzyme and its corresponding pocket mask, and the target reaction data includes the substrate molecular structure data and product molecular structure data of the target chemical reaction; The query data is input into the pedestal prediction model to obtain query hyperbolic representation data; wherein, the pedestal prediction model is obtained by the protease data prediction method for catalytic reactions according to any one of claims 1 to 6; when the query data is target enzyme data, the query hyperbolic representation data is enzyme hyperbolic representation data, including protein hyperbolic representation data and pocket hyperbolic representation data; when the query data is target reaction data, the query hyperbolic representation data is reaction hyperbolic representation data; Obtain a candidate set, which is an enzyme library or a reaction library; wherein, when the query data is target enzyme data, the candidate set is a reaction library, and each candidate reaction in the reaction library corresponds to pre-generated hyperbolic representation data of the reaction; when the query data is target reaction data, the candidate set is an enzyme library, and each candidate enzyme in the enzyme library corresponds to pre-generated hyperbolic representation data of the enzyme. Calculate the similarity between the query hyperbolic representation data and each candidate hyperbolic representation data in the candidate set; Candidate objects are sorted according to the similarity, and at least one candidate object is output based on the sorting results; wherein, when the query data is target enzyme data, the output candidate object is the candidate reaction that the target enzyme can catalyze; when the query data is target reaction data, the output candidate object is the candidate enzyme that catalyzes the target chemical reaction.

10. A method for generating an enzyme protein sequence for a target chemical reaction, characterized in that, Includes the following steps: Obtain molecular structure data of the target chemical reaction, wherein the molecular structure data includes substrate molecular structure data and product molecular structure data; The molecular structure data of the target chemical reaction is input into the pedestal prediction model to obtain hyperbolic representation data of the target reaction; wherein, the pedestal prediction model is obtained by the protease data prediction method for catalytic reactions according to any one of claims 1 to 6; Obtain initial protein sequence data; By iteratively optimizing and adjusting the initial protein sequence data, the difference between the optimized protein sequence data and the target response hyperbolic representation data in the protein pedestal prediction model is reduced; When the preset convergence condition is met, the optimized protein sequence data is output as a candidate enzyme sequence generated for the target chemical reaction.

Images