Methods, devices, and media for designing enzyme sequences and structures using deep learning

By leveraging the synergistic effects of two levels of attention layers—residue-function fusion, spatial neighborhood equivalence, and pocket enhancement—the problem of missing enzyme-substrate interaction modeling and active pocket optimization in enzyme design was solved, enabling efficient and universal enzyme sequence and structure generation to meet diverse industrial catalytic needs.

CN123120035APending Publication Date: 2026-11-17UNIV OF SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611629242.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing deep learning methods in the field of enzyme design suffer from problems such as lack of enzyme-substrate interaction modeling, low accuracy of active pocket optimization, and poor versatility and adaptability, making it difficult to meet the diverse industrial catalysis needs.

Method used

The synergistic effect of residue-function fusion layer, spatial neighborhood isovariance layer and pocket enhancement two-level attention layer is adopted. By modeling heterogeneous spatial graph and multi-scale interaction, enzyme sequence, three-dimensional structure and substrate binding pocket are generated, and supervised training is carried out using EnzyBind dataset.

Benefits of technology

It achieves precise modeling of enzyme-substrate interactions, optimizes the active pocket, and improves substrate binding affinity and catalytic efficiency. It has broad versatility and adaptability, and can meet diverse industrial catalytic needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120035A_ABST
    Figure CN123120035A_ABST
Patent Text Reader

Abstract

This invention belongs to the interdisciplinary field of computer biology, specifically involving a method, device, and medium for designing enzyme sequences and structures using deep learning. This method can jointly generate enzyme sequences, three-dimensional structures, and substrate-binding pockets based on specific substrate conditions. The invention proposes a pocket-enhanced enzyme generation model: generating function-aware residue representations; constructing heterogeneous space maps of enzyme residues and substrate atoms; updating representations and coordinates through type-specific SE(3) isovariant message passing and pocket-aware gating; modeling multi-scale interactions through a two-level attention mechanism consisting of global attention within residues and local attention between residues and substrates; and training the model with a joint objective including losses such as pocket classification, residue-atom contact, interaction distance, and conserved site preservation, ultimately achieving de novo design of substrate-specific enzymes and improving the accuracy and versatility of active pocket optimization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer biology, specifically relating to a method, device, and medium for designing enzyme sequences and structures using deep learning. Background Technology

[0002] Enzymes, as biocatalysts with highly efficient catalytic activity and specificity, have irreplaceable application value in many fields such as industrial manufacturing, drug synthesis, environmental remediation, and food processing. For example, in the pharmaceutical field, enzyme-catalyzed reactions can achieve highly selective conversion of drug precursors and reduce the formation of byproducts; in industrial catalysis, enzymes can replace traditional chemical catalysts, reducing energy consumption and environmental pollution; in environmental remediation, specific enzymes can efficiently degrade pollutants, improving remediation efficiency. However, natural enzymes often suffer from insufficient catalytic activity, poor stability (such as weak thermal stability and pH tolerance), limited substrate specificity, and high production costs, making it difficult to directly meet the needs of industrial applications. Therefore, optimizing the catalytic performance, stability, and specificity of enzymes through artificial design and modification of their amino acid sequences and three-dimensional structures has become a core research direction in the field of enzyme engineering.

[0003] Traditional enzyme design methods are mainly divided into two categories: rational design and directed evolution, but both have significant limitations. Rational design relies on a deep understanding of the relationship between enzyme structure and function, modifying key residues through methods such as site-directed mutagenesis. However, it is heavily reliant on the structural biology knowledge and experience of experimenters, making it difficult to handle the complex nonlinear dependencies between enzyme sequence, structure, and function. Furthermore, its efficiency for multi-residue synergistic optimization is extremely low, often requiring iterative verification over periods ranging from months to years. Directed evolution, on the other hand, constructs enzyme mutant libraries through random mutations and combines this with high-throughput screening to obtain target enzymes. However, this method is costly, time-consuming, and has limited diversity in mutant libraries, making it difficult to overcome the evolutionary bottlenecks of natural enzymes. It can only perform local optimization within the sequence space of natural enzymes and cannot achieve de novo enzyme design.

[0004] In recent years, deep learning technology has made groundbreaking progress in the fields of protein structure prediction and sequence design, providing new technical pathways for enzyme design. For example, deep learning-based protein structure prediction models (such as the AlphaFold series) can accurately predict the three-dimensional structure of proteins based on amino acid sequences, providing an efficient tool for enzyme structure analysis; protein language models (such as the ESM series) achieve sequence design based on structure constraints by learning the evolutionary information of massive amounts of natural protein sequences.

[0005] Inspired by the ESM2 model, the EnzyGen research team proposed a method to address the lack of a unified model and the failure to consider substrate interactions in enzyme design. This method is based on automatically mined functionally important sites and small molecule substrates, jointly generating the enzyme's amino acid sequence and three-dimensional coordinates. The method employs an innovative Neighborhood Attention Equivariance Layer (NAEL) to simultaneously capture long-range sequence associations and the influence of local residues in 3D space. It is also trained using sequence generation loss, position prediction loss, and enzyme-substrate interaction loss. The team also constructed the EnzyBench dataset, containing 3157 enzyme families and 101,974 PDB entries. Experiments showed that EnzyGen performed best among 323 test families, achieving an average substrate binding affinity of -9.44 kcal / mol (10.79% above the best baseline), an ESP score of 0.65 (above the positive binding threshold of 0.6), and an AlphaFold2 pLDDT score of 87.45, demonstrating its ability to design enzymes with stable folding and high substrate affinity.

[0006] However, the application of existing deep learning methods in the field of enzyme design still has many limitations: Existing methods for modeling and design have shortcomings: methods that focus on protein or ligand geometry modeling often fail to accurately capture the interaction between protein and substrate because they neglect to model the interaction. Enzyme-substrate co-design methods, on the other hand, explicitly consider the substrate but neglect modeling the substrate binding pocket, which is fundamental to enzyme-substrate specificity and catalytic activity.

[0007] Low precision in optimizing enzyme activity pockets: The core of enzyme catalytic activity depends on the specific binding of the activity pocket to the substrate and the catalytic reaction, requiring the simultaneous capture of multi-scale non-covalent interactions (such as hydrogen bonds, hydrophobic interactions, and π-π stacking) at the atomic and residue levels. Existing methods struggle to accurately model these multi-scale interactions, resulting in insufficient affinity between the designed activity pocket and the substrate, or limited improvement in catalytic efficiency.

[0008] Poor versatility and adaptability: Existing methods are mostly designed for specific types of enzymes (such as hydrolases and oxidoreductases) or specific substrates. The generalization ability of the models is limited, making it difficult to adapt to diverse industrial catalytic needs. In addition, the generation efficiency is low, making it difficult to produce high-quality enzyme design schemes in batches, which restricts the industrial transformation process. Summary of the Invention

[0009] In view of this, the present invention provides a method, device and medium for designing enzyme sequences and structures using deep learning. It can achieve the joint generation of enzyme sequences, three-dimensional structures and substrate binding pockets by using specific substrates as conditions and through the synergistic effect of residue-function fusion layers, spatial neighborhood isovariance layers and pocket enhancement two-level attention layers. This solves the problems of missing enzyme-substrate interaction modeling, low accuracy of active pocket optimization and poor universality and adaptability of existing methods.

[0010] To achieve the objectives of this invention, the following technical solutions are provided.

[0011] A method for designing enzyme sequences and structures using deep learning includes: Obtain the target enzyme's EC number, the atomic-level chemical characteristics and three-dimensional coordinates of the substrate molecule, and information on functionally conserved sites obtained through multiple sequence alignment; The amino acid type and coordinates of the residues are fused with the functional vector encoded by the EC number to generate a function-aware residue characterization; The enzyme residues and substrate atoms are constructed into a heterogeneous space graph containing three types of edges: residue-residue, atom-atom, and residue-atom. The characterization and coordinates of residues and atoms are updated through type-specific SE(3) isovariant message passing. Multiscale interactions are modeled by a two-level attention mechanism consisting of global attention within residues and local attention between residues and substrates, which updates residue characterization and coordinates and predicts amino acid types. The model was trained using a joint loss function that included sequence reconstruction, coordinate regression, pocket coordinates, binding, pocket classification, residue-atom contact, interaction distance, and conserved site preservation loss. The EC number, substrate information, and functionally conserved sites are input into the trained model to jointly generate the enzyme sequence, three-dimensional structure, and substrate binding pocket structure.

[0012] The method for obtaining the information on the functionally conserved sites includes: obtaining the EC classification of the enzyme from the BRENDA database based on the EC number or predicting the EC classification of the enzyme using the CLEAN model; performing multiple sequence alignment on enzymes of the quaternary EC family using the ClustalW2 tool; and identifying residues that are in the same position in more than 30% of the sequences as functionally conserved sites.

[0013] The substrate binding pocket is a subset of residues in the enzyme whose spatial distance from the substrate is less than a distance threshold of 10 Å. The generation of function-aware residue characterization includes: embedding the amino acid type and coordinates of each residue into an initial residue vector; encoding the EC number into a function vector using a learnable embedding table and adding it to the initial residue vector; then computing the enhanced residue characterization through a Transformer block with a Pre-LN structure using global multi-head self-attention and a feedforward network; and incorporating a geometric attention sublayer in the first Transformer block to fuse geometric information in a rotation- and translation-invariant manner.

[0014] In the heterogeneous spatial diagram: the characteristics of residue nodes include at least amino acid type, function-aware residue characterization, EC function representation, conserved site marker, whether it is located in a candidate pocket, and residue coordinates; the characteristics of substrate atom nodes include at least element type, atomic valence, formal charge, aromaticity, hydrogen bond donor / acceptor properties, hybridization state, and atomic coordinates; residue-substrate atom cross edges are established when the distance between the residue and the substrate atom is less than the pocket neighborhood threshold, and the geometric characteristics of each spatial edge include the radial basis expansion of the distance, relative direction, and optional sequence spacing, and the residue-substrate atom cross edges are also spliced ​​with chemical compatibility features.

[0015] Specifically, independent message functions are used for type-specific message passing for the three types of edges. For residue-substrate atom cross messages, a pocket-aware gating is introduced. The pocket-aware gating is generated by a multilayer perceptron and a sigmoid function from residue representation, substrate atom representation, cross edge features, and the pocket score of the residue or the normalized distance from the substrate. It is used to adjust the magnitude of substrate information injection into residue representation and coordinate update.

[0016] Among them, a two-level neighborhood is used for message passing: a short-range residue-atom edge is constructed with a first threshold to describe direct interactions of hydrogen bonds, electrostatics and hydrophobicity; a pocket context edge is constructed with a second threshold greater than the first threshold; the two-level messages are calculated separately and then fused by gating; and the coordinate update is driven by the relative position vector, and the residue coordinates are updated by combining the neighborhood message and the normalized weight to ensure the variability of SE(3).

[0017] In the two-level attention mechanism, intra-residue global attention is calculated based on radial basis function fusion of geometric proximity to calculate inter-residue attention weights and aggregate messages; residue-substrate local attention calculates cross-attention weights between residues and substrate atoms and aggregates interaction messages; the two-level attention messages are fused through a gating mechanism to update residue representations and coordinates, the gating mechanism consisting of pocket score, Z-score normalization, learnable scaling factor and sigmoid function; finally, the residue representation is projected onto the amino acid vocabulary space to obtain amino acid type prediction results.

[0018] The joint loss function includes: sequence reconstruction loss, which optimizes the amino acid type prediction of masked residues based on cross-entropy loss; coordinate regression loss and pocket coordinate loss, which optimize the coordinate prediction of masked residues and real pocket residues based on Huber loss, respectively; pocket classification loss, which predicts whether each residue belongs to a real substrate binding pocket; residue-atom contact loss, which constructs binary labels for residue-substrate atom pairs with a distance less than the contact threshold and trains them using weighted binary cross-entropy; interaction distance loss, which constrains the distance matrix between real and predicted residues and atoms; and conserved site preservation loss, which only constrains the amino acid type or coordinates to not deviate from the given functional backbone at functionally conserved sites. The model is trained using the EnzyBind enzyme-pocket refining dataset, which integrates BRENDA, PDBBind, and CrossDocked data.

[0019] The present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method for designing enzyme sequences and structures using deep learning as described in the present invention.

[0020] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for designing enzyme sequences and structures using deep learning as described in the present invention.

[0021] Beneficial effects 1. Compared to the process of first screening a small number of key sites based on conformational coordinate differences and then combining mutations, this invention constructs a heterogeneous graph of all candidate pocket residues and substrate atoms within the network, and continuously updates the residue representation and coordinates through learnable type-specific message passing, so that substrate conditions can directly participate in sequence and structure generation, realizing enzyme-pocket-substrate collaborative modeling.

[0022] 2. The spatial neighborhood isovariant layer of the present invention clearly distinguishes three types of relationships: residue-residue, atom-atom, and residue-atom, and uses relative position vectors for isovariant coordinate updates; therefore, the model maintains geometric consistency when rotating or translating the input complex as a whole, and does not confuse the internal structural relationships of the enzyme with the non-covalent relationships between the enzyme and the substrate into the same type of information, thus improving the accuracy of geometric modeling.

[0023] 3. The residue-substrate cross edges of the present invention simultaneously encode spatial distance and substrate atomic chemical properties, and are gated by pocket score; the near neighborhood is used for direct contact, and the larger neighborhood is used for the pocket environment. Mechanistically, this supports differentiated modeling of different local environments such as hydrogen bonding, electrostatic, hydrophobic and aromatic interactions, thereby accurately optimizing the active pocket and improving substrate binding affinity and catalytic efficiency.

[0024] 4. The joint loss of this invention no longer only optimizes the sequence and coordinates, but also supervises the pocket residues, residue-atom contact relationships, interaction distances and functionally conserved sites respectively, so that each key intermediate variable of the model has a corresponding trainable target. This helps to reduce the weakening of substrate information caused by relying solely on the final sequence / coordinate loss, and makes "which residues constitute pockets and which residue-atoms form direct contacts" the training target at the same time.

[0025] 5. The EnzyBind data constructed in this invention organizes enzyme function annotation, enzyme structure, substrate atoms and pocket relationships into a unified sample, providing paired supervision for heterogeneous space graphs and multi-task loss. This data organization method directly corresponds to the residue-substrate cross edges, pocket labels and contact labels in the model, filling the dataset gap for joint modeling of enzymes and binding pockets.

[0026] 6. This invention is not designed for specific types of enzymes or specific substrates. The model can produce high-quality enzyme design schemes in batches using EC numbers and substrate information as conditions. It has broad versatility and adaptability, and can adapt to diverse industrial catalysis needs. It is of great significance for promoting the industrial application of enzymes, reducing production costs, and expanding application scenarios. Attached Figure Description

[0027] Figure 1 This is an overall framework diagram of the PEGM model in this embodiment of the invention. It intuitively shows the complete process from input data (conserved site enzyme sequence and coordinates, EC number, substrate characteristics and coordinates) to processing by the RFF, SNE, and RBA modules, and finally outputting the enzyme sequence, enzyme structure, and substrate binding pocket. It clearly presents the logical relationship and data flow of each module.

[0028] Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0030] This invention provides a deep learning method capable of co-designing enzyme sequences and structures, precisely optimizing active pockets, and possessing broad applicability. This addresses a pressing technical problem in the field of enzyme engineering and is of great significance for promoting the industrial application of enzymes, reducing production costs, and expanding application scenarios. This invention achieves message transmission that distinguishes between the enzyme functional properties of residues and the chemical properties of substrate atoms within the same rotational / translational isovariable space. Furthermore, it makes "which residues constitute pockets and which residue-atoms form direct contacts" simultaneously training objectives, rather than simply selecting key sites based on conformational coordinate differences and then manually combining them.

[0031] This invention proposes a pocket-augmented enzyme generative model (PEGM), which, under specific substrate conditions, achieves the joint generation of enzyme sequence, three-dimensional structure, and substrate binding pocket through the synergistic effect of a residue-function fusion layer, a spatial neighborhood isovariance layer, and a two-tiered attention layer for pocket enhancement. The method flow of this invention is shown in Figure 1, and the specific steps are as follows: Step 1: Input Data Preparation and Preprocessing We collect core enzyme-related data, including the enzyme's amino acid sequence, Cα atom three-dimensional coordinates, Enzyme Committee (EC) number, atomic-level chemical characteristics and three-dimensional coordinates of the substrate molecule, and obtain information on functionally conserved sites through multiple sequence alignment (MSA).

[0032] Enzyme sequence representation: Let A be the set of 20 amino acids, and let N residues be the enzyme sequence. ,in , It is an amino acid type. The coordinates of the Cα atom are three-dimensional. Substrate representation: Substrate containing M atoms ,in a j It has atomic-level chemical characteristics. The coordinates are atomic three-dimensional coordinates; functionally conserved site identification: based on EC numbers obtained from the Brenda database or predicted using the CLEAN model, enzyme EC classification is performed using ClustalW2 for enzymes of the quaternary EC family. Residues that are congruent in more than 30% of the sequence are defined as functionally conserved sites. The conserved site index set is denoted as M, and the corresponding residue set is denoted as M. ; Symbol explanation: The number of enzyme residues. The number of substrate atoms. , These are the residue and substrate atom indices, respectively. Indicates the first Each substrate atom has the following chemical characteristics: 3D coordinates are The set of functionally conserved site indices and the number of substrate atoms. The meanings are different; the former is determined by sequence alignment and cannot be confused with the number of atoms.

[0033] Combined pocket definition: base combined pocket This is the subset of residues in the enzyme whose spatial distance from the substrate satisfies a threshold, i.e.:

[0034] in denoted by L2 norm, and d is the distance threshold (unit: Å), which is actually 10 Å.

[0035] In the formula, Indicates the set of bound pocket residues. This represents the spatial distance threshold between the residue and the substrate atom, measured in angstroms (Å). In this implementation, 10 Å is used; the distance is calculated using the Euclidean norm. This represents spatial distance, and in the scaling attention formula described later, it represents the feature dimension. These are different local meanings.

[0036] Step 2: Residue-Function Fusion (RFF) Module Processing By fusing residue information with prior enzyme function, a function-aware residue characterization is generated: Initial embedding: assigning the amino acid type of each residue and coordinates Embedded as an initial residue vector , For the embedded dimension; EC Number Embedding: EC numbers are embedded through a learnable embedding table. Encoded as a function vector and with the initial residue vector The addition of these components allows for the fusion of enzyme family characteristics at the residue level. In the formula, Here, `Emb` represents the EC number of the target enzyme, and `Emb` is a learnable EC number embedding map. For the first The initial or current feature vector of each residue. The dimension of the residue embedding feature; the feature dimension belongs to the model structure parameters.

[0037] Transformer block processing: A Pre-LN structured Transformer block is used, employing global multi-head self-attention (MHA) to capture long-range dependencies between residues, followed by residue characterization enhanced by a feedforward network (FFN). The update formula for each Transformer block is:

[0038] in For layer normalization operators, It is a set of residue representations output from the previous layer, and incorporates a geometric attention sublayer in the first Transformer block to fuse geometric information in a rotation- and translation-invariant manner.

[0039] Figure 1 In This indicates the number of repetitions of the Transformer block contained in the residue-function fusion module; this section... Indicates the Transformer block layer index. This represents the set of residue features for that layer. LN, MHA, and FFN represent layer normalization, multi-head attention, and feedforward neural network, respectively. The specific value needs to be determined according to the configuration of the model structure.

[0040] Step 3: Spatial Neighborhood Equivariance (SNE) Layer Processing Based on the isovariability of SE (3), geometric reasoning is used to optimize the characterization and coordinates of residues and substrate atoms: Neighborhood message passing: for the first generation of RFF layer output Layer residue characterization and coordinates Initial atomic characterization of the substrate and coordinates Message passing occurs within a local neighborhood of a preset radius, and messages between residues and between atoms are defined as follows: and ( (The number of SNE layer iterations). To achieve deep coupling between the SNE layer and specific enzyme and substrate characteristics, this embodiment constructs a heterogeneous space map of enzyme residues and substrate atoms. .in For enzyme residue nodes, node features include at least amino acid type, RFF output, EC function representation, conserved / motif marker, whether it is located in a candidate pocket, and residue coordinates; For substrate atom nodes, node characteristics include at least element type, atomic valence, formal charge, aromaticity, hydrogen bond donor / acceptor properties, hybridization state, and atomic coordinates.

[0041] Heterogeneous edges are created according to node type: Residue-residue adjacency edges can be established based on Cα distance or local k-nearest neighbors; It represents the chemical bonds or spatial nearest neighbors within the substrate molecule and encodes the bond type; Residue-substrate atom crosses are established when the distance between the residue-representing atom or side-chain atom and the substrate atom is less than the pocket neighborhood threshold. The geometric features of each spatial edge include the radial basis expansion of the distance, the relative direction, and the optional sequence spacing. Residue-atom crosses also incorporate chemical compatibility features such as residue side-chain class, atom donor / acceptor, charge, and aromaticity.

[0042] Use independent message functions for different edge types. Residue-to-residue message. Substrate atom-atomic message Residue-substrate cross-message .in Represented by residues, Let be the substrate atom representation, and 'e' be the corresponding edge feature. By using type-specific parameters, we can avoid treating the covalent / spatial relationships within the enzyme and the non-covalent relationships between the enzyme and substrate as the same interaction.

[0043] For residue-substrate cross messages, a pocket-aware gating system is introduced. ,in The pocket score or normalized distance from the substrate is given to the residue. This is used to adjust the magnitude of substrate information injection residue representation and coordinate updates, enabling residues that are close to the substrate and have matching chemical properties to receive stronger interactive updates.

[0044] In the formula, , and These correspond to residue-residue, substrate-atom, and residue-substrate-atom messages, respectively; their superscripts RR, LL, and RL are used to distinguish the three types of edges. , , These represent the learnable message functions for the corresponding edge types. Indicates the features of intersecting edges; For residues With substrate atoms The gating coefficients between them The pocket score or normalized distance for this residue is σ, where σ represents the sigmoid function and MLP represents the multilayer perceptron.

[0045] Characterization Update: Updating residue and atom characterizations by aggregating neighborhood messages:

[0046]

[0047] in , Residues ,atom The local neighborhood; Coordinate update: The residue coordinates are updated by combining neighborhood messages and relative position vectors to ensure SE (3) isovariability.

[0048] in This represents the number of nodes in the neighborhood, used for normalization. In actual deployment, one SNE layer is stacked every 11 RFF layers to output a geometry-aware residue representation. Updated coordinates and atomic characterization .

[0049] The coordinate updates of the SNE are driven by relative position vectors to ensure synchronized output transformations during global rotation and translation. For residue i, the updates from residue neighbors and substrate atoms can be written as follows: , and then ;in To normalize the weights, The coordinates are those of the substrate atoms. The substrate coordinates can be used as a fixed condition or updated jointly with small amplitudes using the same isovariant rules.

[0050] To simultaneously model direct interactions and a large-scale pocket environment, a two-level neighborhood can be used: a first threshold... Used to construct short-range residue-atom edges, focusing on describing direct interactions such as hydrogen bonds, electrostatics, and hydrophobicity; second threshold ( Furthermore, the 10 Å pocket range from the original scheme can be used to construct pocket context edges. After the two levels of messages are calculated separately, they are fused through gating so that the SNE output simultaneously contains substrate atomic-level local interactions and pocket residue-level environmental information.

[0051] Figure 1 In Indicates the number of repetitions of the spatial neighborhood isovariant layer; , The two-level edge radii are respectively for near-field interaction and pocket context, and satisfy the following conditions: Greater than The second threshold may be a 10 Å pocket range, and the fixed value of the first threshold was not specified in the original application. This indicates the normalized weights used for neighborhood coordinate updates. , These represent the neighborhood index sets of the corresponding nodes. This represents a learnable function used for coordinate updates.

[0052] Step 4: Residue-Atom Two-Level Attention (RBA) Module Processing Multi-scale interactions are modeled using a two-level attention mechanism to enhance pocket features: Intra-residue attention (global): Based on radial basis functions By fusing geometric proximity, calculating inter-residue attention weights, and aggregating messages:

[0053]

[0054] in , , These are respectively residue feature lookup, key-value projection, and value projection. For linear projection weights, For feature dimensions; In the formula, , , This section represents the query vector, bond vector, and value vector for residue attention (different from the input substrate atom set). ); For the corresponding linear projection parameters, Represents the attention weights between residues. This is the result of weighted aggregation; This section describes the dimension of the attention feature. It is a function that converts spatial distance into radial basis features.

[0055] Residue-Substrate Attention (Local): Calculates the cross-attention weights between residues and substrate atoms and aggregates interaction messages.

[0056]

[0057] in , , These are queries for residue and atomic features, and bond and value projections, respectively. For linear projection weights; In the formula, , , This section represents the query, key, and value vectors for residue-substrate cross-attention, respectively. This represents the corresponding learnable projection parameters. Represents residues For substrate atoms Attention weights This indicates the residue-substrate interaction characteristics obtained from polymerization; compared with intra-residue attention, the bond and value in this segment come from the substrate atom.

[0058] Feature and coordinate update: Residue representation and coordinates are updated by fusing two levels of attention messages through a gating mechanism.

[0059]

[0060] in Score for the pocket. For Z-score normalization, For learnable scaling factor, For the sigmoid function, A gating term for controlling the magnitude of coordinate updates; In the formula, This section is used to represent residues. The pocket score has a different meaning from the amino acid type symbol of the same name in the input data definition; Norm is the Z-score normalization operator. Let σ be a learnable scaling factor, and σ be the sigmoid function. A gating term to control the magnitude of coordinate updates, The learnable projection parameters for this gating. Figure 1 In The number of repetitions represents the number of times the residue-atom dual-scale attention module is repeated; the above-mentioned learnable parameters and number of repetitions were not specified as uniform fixed values ​​in the original application.

[0061] Amino acid type prediction: Characterizing the final residues Projecting onto the amino acid vocabulary space yields the prediction results:

[0062] in , For trainable parameters, This refers to the vocabulary of amino acids.

[0063] Step 5: Model Training and Optimization The model is trained using a joint loss function to optimize the generation of enzyme sequences, structures, and binding pockets.

[0064] Sequence reconstruction loss Based on cross-entropy loss ( Optimize the prediction of amino acid types for masked residues:

[0065] Coordinate Regression Loss Based on Huber loss ( Optimize the coordinate prediction of mask residues:

[0066] Pocket coordinate loss : Optimizing the coordinate prediction of real pocket residues based on Huber loss:

[0067] Combined with losses : Optimize the prediction of enzyme-substrate binding sites; in This is the set of masked residues (non-functional conserved sites) used in training. For a true pocket residue set, , Residues The true amino acid type and three-dimensional coordinates, , , , To balance the hyperparameters of various losses.

[0068] in, , , and These are the weights for sequence reconstruction, residue coordinate regression, pocket coordinate regression, and binding term loss, respectively, and are hyperparameters of the training objective; the original application did not specify their fixed values. The masked residue set does not contain pre-specified functionally conserved sites to be protected, and the true pocket set is used for pocket coordinate supervision; cross-entropy loss is used for classification, and Huber loss is used for continuous coordinate regression.

[0069] In addition to the basic loss mentioned above, to ensure that the training objective directly constrains the interaction relationship between the enzyme, pocket, and substrate, the following auxiliary loss is further added: pocket classification loss. This is used to predict whether each residue belongs to a true substrate binding pocket; residue-atom contact loss. Binary labels are constructed for residue-substrate atom pairs in the real structure whose distance is less than the contact threshold, and weighted BCE training is used; interaction distance loss is employed. Huber or smoothed L1 constraints are applied to the actual residue-atom distance matrix and the predicted distance matrix within the pocket; conserved site preservation loss. Only at functionally conserved sites are the amino acid types or coordinates restricted to not deviate from the given functional backbone.

[0070] Therefore, the joint objective can be written as If the training samples also have enzyme-substrate affinity or ranking tags, affinity regression / ranking loss can be added. For samples lacking a certain type of label, a corresponding mask is used to exclude them from that loss term. This design establishes a direct correspondence between the loss function and residue type, pocket location, substrate atom contact, and functionally conserved site, respectively.

[0071] Extended Joint Objectives , , and These are the weights for pocket classification, residue-atom contact, interaction distance, and conserved site retention loss, respectively; these coefficients, together with the aforementioned four coefficients, determine the relative contribution of each supervision term. This is an optional supervision loss when affinity or ranking labels are present.

[0072] Step 6: Output Generation After the model training is completed, the target EC number, substrate information and functionally conserved sites are input. After processing through the above layers, the complete enzyme sequence, three-dimensional structure and corresponding substrate binding pocket structure are output, realizing the de novo design of substrate-specific enzymes.

[0073] Further, the specific implementation details of model training in step 5 are as follows: The training data uses the EnzyBind dataset, which integrates data from BRENDA, PDBBind, and CrossDocked. This dataset organizes enzyme functional annotations, enzyme structures, substrate atoms, and pocket relationships into a unified sample, providing paired supervision for heterogeneous spatial graphs and multi-task losses, filling the dataset gap for joint modeling of enzymes and binding pockets. During training, residues at non-functionally conserved sites are randomly masked. The model reconstructs the amino acid type and coordinates of the masked residues based on contextual residues, EC functional priors, and substrate conditions. If the training samples also have enzyme-substrate affinity or ordination tags, affinity regression / ordination loss can be added. Samples lacking a certain type of tag are masked accordingly, preventing them from participating in that loss. This design establishes a direct correspondence between the loss function and residue type, pocket location, substrate atom contact, and functionally conserved sites.

[0074] Furthermore, the specific implementation details of the inference generation stage in step 6 are as follows: no masking operation is required during inference. The user only needs to provide the target EC number (or its prediction result), the atomic-level chemical characteristics and three-dimensional coordinates of the target substrate molecule, and the functionally conserved sites to be retained. The model will then output the complete enzyme sequence, the three-dimensional coordinates of each residue Cα, and the substrate binding pocket structure in one go. The generated enzyme sequence and structure can be further verified for folding quality through protein structure prediction models (such as AlphaFold2), and substrate binding affinity can be evaluated through molecular docking. This allows for the screening of candidate enzymes with stable folding and high substrate affinity, enabling the de novo design and mass production of substrate-specific enzymes.

[0075] Optionally, in other embodiments: the equivariant message passing of the spatial neighborhood equivariant layer can adopt EGNN, GVP, SE(3)-Transformer or other new networks constructed from relative coordinates; residue nodes can use Cα, main chain atoms or side chain centers as representative coordinates, and substrate nodes can be represented at the atomic or group level; residue-substrate cross-neighborhoods can adopt fixed radius, k-nearest neighbor or a combination of both, and multi-scale neighborhoods can use two or more distance thresholds; pocket sensing gating can be generated by distance, pocket probability, atomic chemical compatibility or attention weights; the pocket, contact, interaction distance, affinity and conserved site terms in the joint loss can be selected and enabled according to the availability of data tags; as long as the heterogeneous spatial interaction between enzyme residues and substrate nodes is still the core, and their pocket / contact relationships are constrained by corresponding supervision, it is an alternative implementation of this technical solution.

[0076] This application also provides an electronic device. Figure 2The structure of an electronic device provided in an embodiment of the present invention is illustrated. For example, the electronic device 20 may include a processor 21, a memory 22, and a transmission device 23. The processor 21 is used to execute the method for designing enzyme sequences and structures using deep learning mentioned in the above embodiments. The processor and the memory can be connected via a bus or other means, taking a bus connection as an example. The transmission device can be connected to the processor and the memory via wired or wireless means. The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the method for designing enzyme sequences and structures using deep learning in the embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the method for designing enzyme sequences and structures using deep learning in the above method embodiments. The memory may include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created by the processor, etc. In addition, the memory may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, and this remote memory may be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The one or more modules stored in the memory, when executed by the processor, perform the method for designing enzyme sequences and structures using deep learning in the embodiments.

[0077] In another aspect, this application also provides a computer-readable storage medium, which may be the computer-readable storage medium included in the apparatus described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into the device. The computer-readable storage medium may be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art. The computer-readable storage medium stores one or more programs, which are used by one or more processors to execute the method described in this application for designing enzyme sequences and structures using deep learning.

[0078] This invention includes, but is not limited to, the above embodiments. Any equivalent substitutions or partial improvements made under the spirit and principles of this invention shall be considered within the scope of protection of this invention.

Claims

1. A method for designing enzyme sequences and structures using deep learning, characterized in that, include: Obtain the target enzyme's EC number, the atomic-level chemical characteristics and three-dimensional coordinates of the substrate molecule, and information on functionally conserved sites obtained through multiple sequence alignment; The amino acid type and coordinates of the residues are fused with the functional vector encoded by the EC number to generate a function-aware residue characterization; The enzyme residues and substrate atoms are constructed into a heterogeneous space graph containing three types of edges: residue-residue, atom-atom, and residue-atom. The characterization and coordinates of residues and atoms are updated through type-specific SE(3) isovariant message passing. Multiscale interactions are modeled by a two-level attention mechanism consisting of global attention within residues and local attention between residues and substrates, which updates residue characterization and coordinates and predicts amino acid types. The model was trained using a joint loss function that included sequence reconstruction, coordinate regression, pocket coordinates, binding, pocket classification, residue-atom contact, interaction distance, and conserved site preservation loss. The EC number, substrate information, and functionally conserved sites are input into the trained model to jointly generate the enzyme sequence, three-dimensional structure, and substrate binding pocket structure.

2. The method according to claim 1, characterized in that, The methods for obtaining the information on the functionally conserved sites include: obtaining the EC classification of the enzyme from the BRENDA database based on the EC number or predicting the EC classification of the enzyme using the CLEAN model; performing multiple sequence alignment on enzymes of the quaternary EC family using the ClustalW2 tool; and identifying residues that are in the same position in more than 30% of the sequences as functionally conserved sites.

3. The method according to claim 2, characterized in that, The substrate binding pocket is a subset of residues in the enzyme whose spatial distance from the substrate is less than a distance threshold, where the distance threshold is 10 Å. The generation of function-aware residue characterization includes: embedding the amino acid type and coordinates of each residue into an initial residue vector; encoding the EC number into a function vector using a learnable embedding table and adding it to the initial residue vector; then computing the enhanced residue characterization through a Pre-LN structure Transformer block using global multi-head self-attention and a feedforward network; and incorporating a geometric attention sublayer in the first Transformer block to fuse geometric information in a rotation- and translation-invariant manner.

4. The method according to claim 3, characterized in that, In the heterogeneous spatial diagram: the characteristics of residue nodes include at least amino acid type, function-aware residue characterization, EC function representation, conserved site marker, whether it is located in a candidate pocket, and residue coordinates; the characteristics of substrate atom nodes include at least element type, atomic valence, formal charge, aromaticity, hydrogen bond donor / acceptor properties, hybridization state, and atomic coordinates; residue-substrate atom cross edges are established when the distance between the residue and the substrate atom is less than the pocket neighborhood threshold, and the geometric characteristics of each spatial edge include the radial basis expansion of the distance, relative direction, and optional sequence spacing, and the residue-substrate atom cross edges are also spliced ​​with chemical compatibility features.

5. The method according to claim 4, characterized in that, Independent message functions are used for type-specific message passing for the three types of edges. For residue-substrate atom cross messages, a pocket-aware gating is introduced. The pocket-aware gating is generated by a multilayer perceptron and a sigmoid function from residue representation, substrate atom representation, cross edge features, and the pocket score of the residue or the normalized distance from the substrate. It is used to adjust the magnitude of substrate information injection into residue representation and coordinate update.

6. The method according to any one of claims 1-5, characterized in that, Two-level neighborhood is used for message passing: a first threshold is used to construct short-range residue-atom edges to describe direct interactions of hydrogen bonds, electrostatics and hydrophobicity; a second threshold greater than the first threshold is used to construct pocket context edges; the two-level messages are calculated separately and then fused by gating; and the coordinate update is driven by the relative position vector, and the residue coordinates are updated by combining neighborhood messages and normalized weights to ensure the variability of SE(3).

7. The method according to any one of claims 1-5, characterized in that, In the two-level attention mechanism, intra-residue global attention is calculated based on radial basis function fusion of geometric proximity to calculate inter-residue attention weights and aggregate messages; residue-substrate local attention calculates cross-attention weights between residues and substrate atoms and aggregates interaction messages; the two-level attention messages are fused through a gating mechanism to update residue representations and coordinates, the gating mechanism consisting of pocket score, Z-score normalization, learnable scaling factor and sigmoid function; finally, the residue representation is projected onto the amino acid vocabulary space to obtain amino acid type prediction results.

8. The method according to any one of claims 1-5, characterized in that, In the joint loss function: sequence reconstruction loss optimizes the amino acid type prediction of mask residues based on cross-entropy loss; coordinate regression loss and pocket coordinate loss optimize the coordinate prediction of mask residues and real pocket residues based on Huber loss, respectively; pocket classification loss is used to predict whether each residue belongs to a real substrate binding pocket; residue-atom contact loss constructs binary labels for residue-substrate atom pairs with a distance less than the contact threshold and trains them using weighted binary cross-entropy; interaction distance loss constrains the distance matrix between real and predicted residue-atom pairs; conserved site preservation loss only constrains the amino acid type or coordinates to not deviate from the given functional backbone at functionally conserved sites; the model is trained using the EnzyBind enzyme-pocket refining dataset, which integrates BRENDA, PDBBind, and CrossDocked data.

9. An electronic device, characterized in that, include: At least one processor; The at least one processor is also connected in communication with a memory, wherein the memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of designing enzyme sequences and structures using deep learning as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method for designing enzyme sequences and structures using deep learning as described in any one of claims 1 to 8.