Multi-modal enzyme kinetic parameter prediction method of adaptive protein language model

Abstract

The invention relates to the cross technical field of artificial intelligence and bioinformatics, in particular to a multi-modal enzyme kinetic parameter prediction method of an adaptive protein language model. Aiming at the problems that an existing processing method is lack of enzymatic reaction dynamic mechanism modeling and insufficient in three-dimensional structure utilization, the invention provides the following technical scheme: step 1, acquiring and preprocessing multi-modal data; 2, protein language model and molecular fingerprint feature extraction; 3, carrying out substrate recognition feature fusion based on cross attention; step 4, extracting conformation adaptive features based on the hybrid expert network; 5, correcting enzyme-substrate distribution alignment characteristics; step 6, kinetic parameter regression prediction; step 7, constructing a multi-objective loss function; and step 8, model training and parameter optimization. Through an enzyme reaction bridging adapter and an enzyme-substrate distribution alignment technology, high-precision prediction of enzyme kinetic parameters is realized, and cross-data-set accuracy and robustness are remarkably improved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of artificial intelligence and bioinformatics, and in particular to a method for predicting multimodal enzyme kinetic parameters adapted to protein language models. Background Technology

[0002] Enzymes, as core functional components in biocatalytic processes, directly quantify catalytic efficiency and substrate affinity through their kinetic parameters (such as transition number kcat, Michaelis constant Km, and inhibition constant Ki). These parameters serve as crucial foundational data for synthetic biology component screening, metabolic pathway optimization design, and targeted drug development. Traditional experimental methods for determining enzyme kinetic parameters are not only time-consuming, labor-intensive, and costly, but also fail to meet the demands of high-throughput screening applications. With the deep penetration of artificial intelligence technology into computational biology, deep learning-based enzyme kinetic parameter prediction models have become an important research direction replacing traditional experimental methods, significantly improving parameter acquisition efficiency.

[0003] In recent years, enzyme kinetic parameter prediction methods have undergone multiple rounds of technological iteration, gradually evolving from early shallow machine learning models to deep learning architectures that integrate multimodal information. Three main technical paths have emerged: First, shallow fusion architectures based on sequence and molecular structure. While these methods achieve a preliminary combination of multimodal information, the depth of feature interaction is insufficient. Second, prediction architectures based on pre-trained protein language models (PLMs) acquire deep evolutionary and biochemical semantic features of enzyme sequences through feature freezing or fine-tuning, significantly improving prediction accuracy and becoming the current mainstream technical solution. Third, geometric enhancement architectures that introduce three-dimensional structural information extract three-dimensional geometric topological features of enzyme active sites through models such as isomorphic graph neural networks, attempting to compensate for the lack of spatial structural information in one-dimensional sequence representation, and hoping to further improve the model's ability to model enzyme-catalyzed reaction mechanisms.

[0004] Despite the progress made in existing technologies, significant limitations remain in simulating real enzymatic reactions, extracting the value of geometric structures, and ensuring the stability of multimodal fusion. These limitations severely restrict the predictive accuracy and generalization ability of the models. Specifically, the lack of modeling for the dynamic mechanisms of enzymatic reactions at different stages is a major issue. Real enzymatic reactions involve a dynamic process of "substrate-specific recognition - active site conformational adaptation - catalytic reaction occurrence." However, existing multimodal prediction models generally employ static fusion strategies. Whether through simple feature splicing or single cross-attention interaction, neither differentiates the modeling of the two key stages of "substrate recognition" and "conformational adaptation." This static modeling approach fails to capture the dynamic conformational changes of the enzyme's active site after substrate binding, making it difficult for the model to accurately characterize the specific interaction mechanism between the enzyme and the substrate, resulting in deviations between the predicted results and the actual enzymatic reaction patterns.

[0005] Insufficient utilization of the geometric heterogeneity of active sites: Enzyme active sites exhibit significant structural diversity and conformational flexibility. When binding to different substrates, the active site pockets undergo specific local conformational adjustments to adapt to the substrate structure. Existing methods for introducing 3D structural information mostly employ a single shared network module to process the geometric features of all active sites, failing to design differentiated feature extraction mechanisms for pocket structures with different geometric topologies. In particular, there is a lack of adaptive modeling strategies that can dynamically match pocket conformational changes, failing to fully explore the regulatory value of geometric structural information on enzyme-catalyzed reactions. Consequently, the introduction of 3D structural information fails to achieve the expected performance improvement, and in some scenarios, feature redundancy even reduces model generalization.

[0006] Lack of distributional stability in multimodal feature fusion: Existing methods generally lack effective distributional alignment constraint mechanisms when fusing sequential semantic features and 3D geometric features from pre-trained protein language models. Since sequential semantic features and geometric features originate from different data distribution spaces, direct forced fusion easily leads to semantic conflicts or one feature being overwhelmed by the other, disrupting the biochemical semantic manifold originally learned by the protein language model and causing a "catastrophic forgetting" of prior biochemical knowledge. Furthermore, during model training, if complex architectures such as multi-expert models are used, uneven utilization of expert networks often occurs, further exacerbating the instability of the feature space. This results in extremely poor robustness of the model in prediction tasks involving cross-datasets and out-of-distribution samples, making it difficult to meet practical application requirements.

[0007] Furthermore, existing models that rely on multi-expert networks often fail to design effective load balancing mechanisms, resulting in some expert networks being idle for extended periods while others are overloaded, thus failing to fully leverage the adaptability advantages of multi-expert architectures to complex heterogeneous features. At the same time, the lack of constraints on the stability of feature space distribution during model training makes it difficult for trained models to maintain stable predictive performance when processing unknown enzyme-related data, limiting their generalization ability.

[0008] In summary, accurately modeling the dynamic process of enzyme-catalyzed reactions, efficiently utilizing the geometric heterogeneity information of active sites, ensuring the distributional stability of multimodal feature fusion, and simultaneously optimizing the utilization balance and feature space stability of expert networks during model training have become core issues that current enzyme kinetic parameter prediction technologies urgently need to address. Therefore, this invention proposes a multimodal enzyme kinetic parameter prediction method adapted to protein language models. Summary of the Invention

[0009] The purpose of this invention is to address the problems in the existing processing methods in the background art that lack modeling of the dynamic mechanism of enzyme-catalyzed reactions and insufficient utilization of three-dimensional structures, and to propose a multimodal enzyme kinetic parameter prediction method adapted to protein language models.

[0010] The technical solution of this invention: a multimodal enzyme kinetic parameter prediction method adapted to protein language models, comprising the following processing steps:

[0011] Step 1: Multimodal data acquisition and preprocessing: Prepare a training dataset containing data from all three modalities, clean and standardize the dataset to ensure that the information from the three modalities corresponds one-to-one;

[0012] Step 2, Protein Language Model and Molecular Fingerprint Feature Extraction: Using a pre-trained protein language model, molecular encoder, and geometric encoder, semantic embedding features of enzyme sequence, chemical features of substrate, and three-dimensional geometric features of active site are extracted, respectively.

[0013] Step 3: Substrate recognition feature fusion based on cross-attention: Input the enzyme sequence semantic embedding features and substrate chemical features into the cross-attention module, inject substrate information to generate conditional features for the substrate recognition stage;

[0014] Step 4: Conformation adaptation feature extraction based on hybrid expert network: Combining substrate recognition conditional features and three-dimensional geometric features of active sites, the hybrid expert network is dynamically activated to generate deep features for the conformation adaptation stage.

[0015] Step 5, Enzyme-Substrate Distribution Alignment Feature Correction: Constrain the feature distribution distance through the enzyme-substrate distribution alignment module to ensure that the multimodal fusion features are consistent with the biochemical semantic manifold of the pre-trained protein language model;

[0016] Step 6: Kinetic parameter regression prediction: The regression prediction head is used to process deep conformational adaptation features to achieve enzyme kinetic parameter regression prediction;

[0017] Step 7: Construction of Multi-Objective Loss Function: Based on the prediction results and the true labels, construct a multi-objective comprehensive loss function that includes regression, expert load balancing, and distribution alignment losses;

[0018] Step 8, Model Training and Parameter Optimization: Adjust the model parameters through backpropagation, and iteratively train the model based on a multi-objective loss function until the performance meets the target, which is then used to predict the kinetic parameters of the enzyme to be tested.

[0019] Optionally, the three-mode data mentioned in step one are strictly paired enzyme amino acid sequences, substrate SMILES molecular formulas, and corresponding three-dimensional structure data of active sites.

[0020] The dataset is cleaned and standardized to ensure that the three modal information of enzyme amino acid sequence, substrate SMILES molecular formula and active site three-dimensional structure are in one-to-one correspondence, thus forming a standardized multimodal dataset for model training.

[0021] Optionally, the molecular encoder in step two is a molecular graph neural network or a fingerprint encoder, and the geometric graph encoder is an E-GNN geometric graph encoder;

[0022] The enzyme amino acid sequence is input into a pre-trained protein language model to extract sequence semantic embedding features. ;

[0023] Input the molecular formula of the substrate SMILES into a molecular graph neural network or fingerprint encoder to extract the chemical characteristics of the substrate. ;

[0024] The three-dimensional pocket data of active sites are input into the geometric graph encoder E-GNN to extract the three-dimensional geometric structure features. ;

[0025] in These represent sequence length, number of substrate atoms, and number of residues, respectively. For feature dimension, It represents the set of real numbers.

[0026] Optionally, the cross-attention module mentioned in step three is a molecular recognition cross-attention module;

[0027] The specific process is as follows:

[0028] Embedding sequence semantics features Chemical characteristics of the substrate The input molecular recognition cross-attention module injects substrate information into the enzyme sequence representation through a cross-attention mechanism, generating conditional features for the substrate recognition stage. Using trainable projection matrices By calculating attention weights and context updates, intermediate features that aggregate substrate information are obtained. , The set of real numbers is represented by the formula:

[0029]

[0030]

[0031] The intermediate features are fused back into the enzyme sequence representation through residual connections and layer normalization to generate conditional features for the substrate recognition stage:

[0032]

[0033] This process mimics the specific recognition of a substrate by an enzyme, allowing the enzyme's characterization to be specifically adjusted according to the chemical properties of the particular substrate.

[0034] Optionally, the hybrid expert network described in step four is a geometry-aware hybrid expert module;

[0035] By pooling active site residues and combining them with conditional features of the substrate recognition stage. Generate geometric routing vectors to guide subsequent expert selection. Specifically, define It is a set of residue indices in the enzyme sequence that belong to the active site pocket;

[0036] Integrate identification and geometric signals using a joint pooling strategy:

[0037] ,

[0038] in, Mean pooling operation representing features This represents vector concatenation, routing vector. It is used to simultaneously reflect the current substrate bonding state and the local pocket geometry and topology, providing a basis for subsequent conformational adaptation decisions.

[0039] Optionally, the enzyme-substrate distribution alignment module described in step five uses a Gaussian kernel function to calculate the maximum mean difference index to constrain the distribution distance, thus constraining the features within the same semantic manifold. Specifically, it defines global feature representations for different stages:

[0040] ,

[0041] in These correspond to the initial sequence, MRCA output, and G-MoE output, respectively; using the Gaussian kernel function. Calculate the maximum mean difference:

[0042]

[0043] By minimizing this distribution alignment loss, we can avoid catastrophic forgetting of biochemical semantics in pre-trained protein language models.

[0044] Optionally, the regression prediction head described in step six is ​​a heteroscedastic Gaussian regression prediction head;

[0045] Deep features of the conformational adaptation stage Input the heteroscedastic Gaussian regression prediction head, and output the corresponding enzyme kinetic parameter prediction results and uncertainty estimates;

[0046] Specifically, the mean of the prediction head output and logarithmic variance :

[0047]

[0048] The predicted values ​​of the target dynamic parameters are: , Used to construct heteroscedastic Gaussian negative log-likelihood loss.

[0049] Optionally, the regression task loss described in step seven is heteroscedastic Gaussian negative log-likelihood loss, and the distribution alignment loss is MMD loss. The three are weighted by preset weights to form a comprehensive loss function.

[0050] Specific total loss function Loss due to regression task Expert load balancing losses and distribution alignment loss composition:

[0051]

[0052] .

[0053] Optionally, the training process in step eight is used to maintain the balance of expert network utilization and the stability of feature space distribution;

[0054] The trained model predicts the parameters to be tested by processing data related to unknown enzymes.

[0055] In summary, this application includes at least one of the following beneficial technical effects:

[0056] This invention proposes an enzyme-reaction bridging adapter, which, through the synergistic effect of a molecular recognition cross-attention module and a geometric perception hybrid expert module, achieves phased conditional modeling of substrate-specific recognition and active site conformational adaptation. This overcomes the limitation of existing technologies that simplify dynamic catalytic processes to static feature splicing. Specifically, the MRCA module simulates the enzyme's specific recognition process of the substrate, enabling enzyme sequence characterization to be specifically adjusted according to the chemical properties of the substrate; the G-MoE module simulates the induced fit mechanism of enzyme-substrate binding, dynamically adapting to the conformational changes of the active site pocket. This invention accurately replicates the core dynamic links of real enzymatic reactions, enabling the model to capture the specific interaction rules between the enzyme and the substrate, significantly improving the consistency between the predicted kinetic parameters and the actual catalytic mechanism.

[0057] This invention constructs a multi-structure expert network through a G-MoE module, combines the three-dimensional geometric features of active sites with substrate recognition features to generate a joint routing vector, and dynamically activates the top-k expert networks that adapt to the current pocket structure for low-rank adaptive transformation. This strategy abandons the drawback of the traditional single shared module processing all geometric structures, can accurately capture the local conformational dynamic changes of the active site pocket when different substrates bind, fully explore the regulatory value of three-dimensional geometric information on enzyme-catalyzed reactions, and effectively solve the problem of poor effect of introducing three-dimensional structures in the prior art, and even reduced generalization due to feature redundancy. In particular, it improves the prediction accuracy for complex scenarios such as highly active mutants.

[0058] Furthermore, by introducing an enzyme-substrate distribution alignment (ESDA) module, the maximum mean difference (MMD) index is used to constrain the distribution distance of sequence semantic features, substrate recognition features, and conformational adaptation features, ensuring that the multimodal fusion features are always kept within the biochemical semantic manifold of the pre-trained protein language model (PLM). This effectively solves the semantic conflict and feature submersion problems caused by the direct fusion of sequence semantic features and three-dimensional geometric features in existing technologies.

[0059] This invention employs a heteroscedastic Gaussian regression prediction head. By outputting the predicted mean and logarithmic variance, it can not only achieve high-precision regression prediction of key kinetic parameters such as enzyme turnover number (kcat), Michaelis constant (Km), and inhibition constant (Ki), but also estimate the uncertainty of the prediction results, providing a reliable confidence reference for subsequent experimental verification. Furthermore, the staged multimodal fusion strategy enables the model to integrate complementary information from enzyme sequence, substrate chemical structure, and the three-dimensional structure of active sites, generating deeper features with greater biochemical interpretability. Compared to existing single-modal or shallow fusion models, both prediction accuracy and generalization ability are significantly improved.

[0060] In summary, this invention utilizes an enzyme reaction bridging adapter and enzyme-substrate distribution alignment technology to model the enzyme-catalyzed dynamic process in stages. Combined with a distribution alignment strategy, it achieves high-precision prediction of enzyme kinetic parameters, significantly improving the accuracy and robustness across datasets. Attached Figure Description

[0061] Figure 1 This is a flowchart of the method proposed in this invention;

[0062] Figure 2 This is a diagram illustrating the entire process from multimodal data input to enzyme kinetic parameter output in this embodiment. Detailed Implementation

[0063] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0064] Example

[0065] Reference Figure 1 and Figure 2 As shown, the multimodal enzyme kinetic parameter prediction method adapted to protein language models proposed in this invention includes the following processing steps:

[0066] Step 1: Multimodal data acquisition and preprocessing: Prepare a training dataset containing data from all three modalities, clean and standardize the dataset to ensure that the information from the three modalities corresponds one-to-one;

[0067] The three-mode data consist of strictly paired enzyme amino acid sequences, substrate SMILES molecular formulas, and corresponding three-dimensional structural data of active sites.

[0068] The dataset was cleaned and standardized to ensure a one-to-one correspondence between the enzyme amino acid sequence, substrate SMILES molecular formula, and active site three-dimensional structure, forming a standardized multimodal dataset for model training; each data point must contain the enzyme sequence. Substrate SMILES string and corresponding three-dimensional pocket structure data of active sites .

[0069] Step 2, Protein Language Model and Molecular Fingerprint Feature Extraction: Using a pre-trained protein language model, molecular encoder, and geometric encoder, semantic embedding features of enzyme sequence, chemical features of substrate, and three-dimensional geometric features of active site are extracted, respectively.

[0070] The molecular encoder is a molecular graph neural network or a fingerprint encoder, and the geometric graph encoder is an E-GNN geometric graph encoder.

[0071] enzyme sequence Input the pre-trained protein language model to extract sequence semantic embedding features. ;

[0072] substrate Input a molecular fingerprint encoder or graph neural network to extract the chemical features of the substrate. ;

[0073] Three-dimensional structure Input geometric graph encoding E-GNN to extract 3D geometric structure features ;

[0074] in Indicates the sequence length. Indicates the number of substrate atoms. Number of residues The dimension of the feature vector. It represents the set of real numbers.

[0075] Step 3: Substrate recognition feature fusion based on cross-attention: Input the enzyme sequence semantic embedding features and substrate chemical features into the cross-attention module, inject substrate information to generate conditional features for the substrate recognition stage;

[0076] The cross-attention module is a molecular recognition cross-attention module;

[0077] The specific process is as follows:

[0078] Embedding sequence semantics features Chemical characteristics of the substrate The input molecular recognition cross-attention module injects substrate information into the enzyme sequence representation through a cross-attention mechanism, generating conditional features for the substrate recognition stage. Using trainable projection matrices By calculating attention weights and context updates, intermediate features that aggregate substrate information are obtained. The specific calculation formula is as follows:

[0079]

[0080]

[0081] In this formula:

[0082] This represents the enzyme-substrate attention weight matrix, characterizing the degree of attention that enzyme residues pay to substrate atoms;

[0083] These represent the trainable Query, Key, and Value projection matrices, respectively.

[0084] The dimension representing the attention head, This is the scaling factor;

[0085] This represents an intermediate feature representation that aggregates substrate information;

[0086] The intermediate features are fused back into the enzyme sequence representation through residual connections and layer normalization to generate conditional features for the substrate recognition stage. :

[0087]

[0088] in, Presentation layer normalization operation.

[0089] This process mimics the specific recognition of a substrate by an enzyme, allowing the enzyme's characterization to be specifically adjusted according to the chemical properties of the particular substrate.

[0090] Step 4: Conformation adaptation feature extraction based on hybrid expert network: Combining substrate recognition conditional features and three-dimensional geometric features of active sites, the hybrid expert network is dynamically activated to generate deep features for the conformation adaptation stage.

[0091] The hybrid expert network is a geometrically perceptive hybrid expert module;

[0092] By pooling active site residues and combining them with conditional features of the substrate recognition stage. Generate geometric routing vectors to guide subsequent expert selection. Specifically, define It is a set of residue indices in the enzyme sequence that belong to the active site pocket;

[0093] Integrate identification and geometric signals using a joint pooling strategy:

[0094] ,

[0095] in, , representing the set of residue indices in the enzyme sequence that belong to the pocket region of the active site;

[0096] This represents the pocket region features extracted from the full sequence features;

[0097] Mean pooling, representing the feature, converts a variable-length sequence into a fixed-length vector;

[0098] This represents a vector concatenation operation;

[0099] Routing vectors It is used to simultaneously reflect the current substrate bonding state and the local pocket geometry and topology, providing a basis for subsequent conformational adaptation decisions.

[0100] Step 5: Enzyme-substrate distribution alignment feature correction: Construct a geometrically perceptual hybrid expert module to extract features from the conformational adaptation stage. ;

[0101] Furthermore, in this step, the routing probability is first calculated using a gated network:

[0102] In this formula:

[0103] Represents the trainable weights and biases of the gating network;

[0104] The total number of representatives in the expert network;

[0105] The activation probability vector representing all experts;

[0106] The operation indicates that only the one with the highest probability is retained. Set one value to zero and the rest to obtain a sparse gated vector. .

[0107] The selected number An expert Perform a low-rank adaptive geometric adjustment transformation:

[0108]

[0109] In this formula:

[0110] Represents the set of real numbers;

[0111] Represents a low-rank projection matrix, where For rank;

[0112] The projection matrix representing geometric features;

[0113] This represents the GELU nonlinear activation function.

[0114] Final conformational adaptation features We obtain this from weighted aggregation:

[0115] in This indicates that index n is taken from the set of the top k most relevant elements selected. Represents a multilayer perceptron. For the first Normalized weights for each expert.

[0116] By constraining the feature distribution distance through the enzyme-substrate distribution alignment module, the multimodal fusion features are made consistent with the biochemical semantic manifold of the pre-trained protein language model;

[0117] The enzyme-substrate distribution alignment module uses a Gaussian kernel function to calculate the maximum mean difference index to constrain the distribution distance, thus binding the features within the same semantic manifold. Specifically, it defines global feature representations for different stages:

[0118] ,

[0119] in Indicates the length of the enzyme's amino acid sequence. This is an index of amino acid residues in the sequence; where These correspond to the initial sequence, MRCA output, and G-MoE output, respectively; using the Gaussian kernel function. Calculate the maximum mean difference:

[0120]

[0121] These represent the initial sequence features respectively. Identification of stage characteristics and adaptation stage characteristics The global mean pooling vector;

[0122] Represents the squared maximum mean difference based on the reproducing kernel Hilbert space (RKHS), used to measure the distance between two distributions;

[0123] By minimizing this distribution alignment loss, we can avoid catastrophic forgetting of biochemical semantics in pre-trained protein language models.

[0124] Step 6: Kinetic parameter regression prediction: The regression prediction head is used to process deep conformational adaptation features to achieve enzyme kinetic parameter regression prediction;

[0125] The regression prediction head is a heteroscedastic Gaussian regression prediction head;

[0126] Deep features of the conformational adaptation stage Input the heteroscedastic Gaussian regression prediction head, and output the corresponding enzyme kinetic parameter prediction results and uncertainty estimates;

[0127] Specifically, the mean of the prediction head output and logarithmic variance :

[0128]

[0129] Among them, This represents a prediction head network consisting of fully connected layers. This is the predicted value of the target dynamic parameters. This represents the uncertainty of the forecast.

[0130] Step 7: Construction of Multi-Objective Loss Function: Based on the prediction results and the true labels, construct a multi-objective comprehensive loss function that includes regression, expert load balancing, and distribution alignment losses;

[0131] The regression task loss is heteroscedastic Gaussian negative log-likelihood loss, and the distribution alignment loss is MMD loss. The three are weighted by preset weights to form a comprehensive loss function.

[0132] Specific total loss function Loss due to regression task Expert load balancing losses and distribution alignment loss composition:

[0133]

[0134] ;

[0135] In this formula: The actual label value; This is a weighting term based on uncertainty, used to dynamically adjust the error penalty intensity;

[0136] , representing the heteroscedastic Gaussian negative log-likelihood loss, is used for supervised regression tasks. This is a real label;

[0137] To compensate for expert load balancing losses, used to prevent the gated network from collapsing to a single expert;

[0138] To balance the hyperparameters of each loss term.

[0139] Step 8, Model Training and Parameter Optimization: Adjust the model parameters through backpropagation, and iteratively train the model based on the multi-objective loss function until the performance meets the target, which is then used for predicting the kinetic parameters of the enzyme to be tested.

[0140] The training process is used to maintain the balance of expert network utilization and the stability of feature space distribution;

[0141] The trained model predicts the parameters to be tested by processing data related to unknown enzymes.

[0142] This embodiment achieves high-precision prediction of key kinetic properties of enzymes, such as the number of enzymes and affinity, by jointly optimizing the above objectives. While ensuring prediction accuracy, the model maintains the balanced utilization of the expert network and the stability of the distribution of the potential feature space.

[0143] The above specific embodiments are merely optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A multimodal enzyme kinetic parameter prediction method adapted to protein language models, characterized in that, The following processing steps are included: Step 1: Multimodal data acquisition and preprocessing: Prepare a training dataset containing data from all three modalities, clean and standardize the dataset to ensure that the information from the three modalities corresponds one-to-one; Step 2, Protein Language Model and Molecular Fingerprint Feature Extraction: Using a pre-trained protein language model, molecular encoder, and geometric encoder, semantic embedding features of enzyme sequence, chemical features of substrate, and three-dimensional geometric features of active site are extracted, respectively. Step 3: Substrate recognition feature fusion based on cross-attention: Input the enzyme sequence semantic embedding features and substrate chemical features into the cross-attention module, inject substrate information to generate conditional features for the substrate recognition stage; Step 4: Conformation adaptation feature extraction based on hybrid expert network: Combining substrate recognition conditional features and three-dimensional geometric features of active sites, the hybrid expert network is dynamically activated to generate deep features for the conformation adaptation stage. Step 5, Enzyme-Substrate Distribution Alignment Feature Correction: Constrain the feature distribution distance through the enzyme-substrate distribution alignment module to ensure that the multimodal fusion features are consistent with the biochemical semantic manifold of the pre-trained protein language model; Step 6: Kinetic parameter regression prediction: The regression prediction head is used to process deep conformational adaptation features to achieve enzyme kinetic parameter regression prediction; Step 7: Construction of Multi-Objective Loss Function: Based on the prediction results and the true labels, construct a multi-objective comprehensive loss function that includes regression, expert load balancing, and distribution alignment losses; Step 8, Model Training and Parameter Optimization: Adjust the model parameters through backpropagation, and iteratively train the model based on a multi-objective loss function until the performance meets the target, which is then used to predict the kinetic parameters of the enzyme to be tested.

2. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model according to claim 1, characterized in that, The three-mode data mentioned in step one are strictly paired enzyme amino acid sequences, substrate SMILES molecular formulas, and corresponding three-dimensional structure data of active sites; The dataset is cleaned and standardized to ensure that the three modal information of enzyme amino acid sequence, substrate SMILES molecular formula and active site three-dimensional structure are in one-to-one correspondence, thus forming a standardized multimodal dataset for model training.

3. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model according to claim 1, characterized in that, The molecular encoder mentioned in step two is a molecular graph neural network or a fingerprint encoder, and the geometric graph encoder is an E-GNN geometric graph encoder; The enzyme amino acid sequence is input into a pre-trained protein language model to extract sequence semantic embedding features. ; Input the molecular formula of the substrate SMILES into a molecular graph neural network or fingerprint encoder to extract the chemical characteristics of the substrate. ; The three-dimensional pocket data of active sites are input into the geometric graph encoder E-GNN to extract the three-dimensional geometric structure features. ; in These represent sequence length, number of substrate atoms, and number of residues, respectively. For feature dimension, It represents the set of real numbers.

4. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model according to claim 1, characterized in that, The cross-attention module mentioned in step three is a molecular recognition cross-attention module; The specific process is as follows: Embedding sequence semantics features Chemical characteristics of the substrate The input molecular recognition cross-attention module injects substrate information into the enzyme sequence representation through a cross-attention mechanism, generating conditional features for the substrate recognition stage. Using trainable projection matrices ,in Representing the set of real numbers, we calculate attention weights and context updates to obtain intermediate features that aggregate substrate information. The formula used is: ， ； The intermediate features are fused back into the enzyme sequence representation through residual connections and layer normalization to generate conditional features for the substrate recognition stage: ， This process mimics the specific recognition of a substrate by an enzyme, allowing the enzyme's characterization to be specifically adjusted according to the chemical properties of the particular substrate.

5. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model according to claim 1, characterized in that, The hybrid expert network described in step four is a geometry-aware hybrid expert module; By pooling active site residues and combining them with conditional features of the substrate recognition stage. Generate geometric routing vectors to guide subsequent expert selection. Specifically, define It is a set of residue indices in the enzyme sequence that belong to the active site pocket; Integrate identification and geometric signals using a joint pooling strategy: ， in, Mean pooling operation representing features This represents vector concatenation, routing vector. It is used to simultaneously reflect the current substrate bonding state and the local pocket geometry and topology, providing a basis for subsequent conformational adaptation decisions.

6. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model according to claim 1, characterized in that, Step five describes an enzyme-substrate distribution alignment module that uses a Gaussian kernel function to calculate the maximum mean difference index to constrain the distribution distance, thus binding the features within the same semantic manifold. Specifically, it defines global feature representations for different stages: ， in These correspond to the initial sequence, MRCA output, and G-MoE output, respectively; using the Gaussian kernel function. Calculate the maximum mean difference: ； By minimizing this distribution alignment loss, we can avoid catastrophic forgetting of biochemical semantics in pre-trained protein language models.

7. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model according to claim 1, characterized in that, The regression prediction head described in step six is ​​a heteroscedastic Gaussian regression prediction head; Deep features of the conformational adaptation stage Input the heteroscedastic Gaussian regression prediction head, and output the corresponding enzyme kinetic parameter prediction results and uncertainty estimates; Specifically, the mean of the prediction head output and logarithmic variance : ， The predicted values ​​of the target dynamic parameters are: , Used to construct heteroscedastic Gaussian negative log-likelihood loss.

8. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model according to claim 1, characterized in that, The regression task loss described in step seven is heteroscedastic Gaussian negative log-likelihood loss, and the distribution alignment loss is MMD loss. The three are weighted by preset weights to form a comprehensive loss function. Specific total loss function Loss due to regression task Expert load balancing losses and distribution alignment loss composition: ， 。 9. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model according to claim 1, characterized in that, The training process in step eight is used to maintain the balance of expert network utilization and the stability of feature space distribution; The trained model predicts the parameters to be tested by processing data related to unknown enzymes.

Images