CN-121983135-A - Multi-modal enzyme dynamics parameter prediction method adapting to protein language model

CN121983135ACN 121983135 ACN121983135 ACN 121983135ACN-121983135-A

Abstract

The invention relates to the technical field of artificial intelligence and bioinformatics intersection, in particular to a multi-modal enzyme dynamics parameter prediction method adapting to a protein language model. The method mainly aims at solving the problems of missing modeling of an enzymatic reaction dynamic mechanism and insufficient utilization of a three-dimensional structure in the existing processing method, and provides the following technical scheme of step one, multi-mode data acquisition and preprocessing, step two, protein language model and molecular fingerprint feature extraction, step three, substrate recognition feature fusion based on cross attention, step four, conformational adaptation feature extraction based on a mixed expert network, step five, enzyme-substrate distribution alignment feature correction, step six, dynamic parameter regression prediction, step seven, multi-objective loss function construction, step eight, model training and parameter optimization. According to the invention, through the technology of aligning the enzyme reaction bridging adapter with the enzyme-substrate distribution, the high-precision prediction of the enzyme kinetic parameters is realized, and the accuracy and the robustness of the cross-data set are obviously improved.

Inventors

WANG FEI
WEI YANYAN
ZHENG XINYE
BAO TONG
YANG JINGWEN

Assignees

合肥工业大学

Dates

Publication Date: 20260505
Application Date: 20260108

Claims (9)

1. The multi-modal enzyme kinetic parameter prediction method adapting to the protein language model is characterized by comprising the following processing steps: step one, multi-modal data acquisition and preprocessing, namely preparing a training data set containing three-modal data, cleaning and standardizing the data set, and ensuring one-to-one correspondence of the three-modal information; Extracting a protein language model and molecular fingerprint characteristics, namely respectively extracting enzyme sequence semantic embedding characteristics, substrate chemical characteristics and active site three-dimensional geometric characteristics through a pre-training protein language model, a molecular encoder and a geometric figure encoder; step three, substrate recognition feature fusion based on cross attention, namely inputting enzyme sequence semantic embedding features and substrate chemical features into a cross attention module, and injecting substrate information to generate substrate recognition stage conditioning features; Step four, extracting conformation adaptive features based on a mixed expert network, namely combining substrate recognition conditional features and active site three-dimensional geometric features, and dynamically activating the expert network through the mixed expert network to generate conformation adaptive stage deep features; step five, correcting the enzyme-substrate distribution alignment characteristic, namely constraining the characteristic distribution distance through an enzyme-substrate distribution alignment module to ensure that the multi-mode fusion characteristic is consistent with the biochemical semantic manifold of the pre-training protein language model; Step six, dynamic parameter regression prediction, namely processing conformation adaptation deep features by using a regression prediction head to realize enzyme dynamic parameter regression prediction; Constructing a multi-objective comprehensive loss function containing regression, expert load balancing and distribution alignment loss based on the prediction result and the real label; and step eight, model training and parameter optimization, namely adjusting model parameters through back propagation, and iteratively training based on a multi-objective loss function until the performance reaches the standard, wherein the model parameters are used for predicting kinetic parameters of the enzyme to be detected.
2. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model according to claim 1, wherein in step one, the three-dimensional data are strictly paired enzyme amino acid sequences, substrate SMILES molecular formulas and corresponding active site three-dimensional structure data; and cleaning and standardizing the data set to ensure that three-mode information of enzyme amino acid sequences, substrate SMILES molecular formulas and active site three-dimensional structures are in one-to-one correspondence, so as to form a standardized multi-mode data set for model training.
3. The method for predicting the dynamic parameters of the multi-modal enzyme adapted to the protein language model according to claim 1, wherein in the second step, the molecular encoder is a molecular map neural network or a fingerprint encoder, and the geometric encoder is an E-GNN geometric encoder; Inputting the enzyme amino acid sequence into a pre-training protein language model, extracting to obtain sequence semantic embedded features ; Inputting the molecular formula of substrate SMILES into a molecular graph neural network or fingerprint encoder, and extracting to obtain chemical characteristics of substrate ; Inputting three-dimensional pocket data of active site into a geometric figure encoder E-GNN, and extracting to obtain three-dimensional geometric structural features ; Wherein the method comprises the steps of Respectively representing the sequence length, the number of substrate atoms and the number of residues, As a dimension of the features, Representing a set of real numbers.
4. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model of claim 1 wherein step three the cross-attention module is a molecular recognition cross-attention module; the specific process is as follows: Embedding sequence semantics into features Chemical characterization with substrate Input molecule recognition cross-attention module, injecting substrate information into enzyme sequence representation through cross-attention mechanism, generating conditional features of substrate recognition stage Using a trainable projection matrix Wherein Representing a real set, computing attention weights and context updates to obtain intermediate features aggregating substrate information The formula is adopted: , ; Fusing the intermediate features back into the enzyme sequence representation by residual ligation and layer normalization, generating a conditional feature of the substrate recognition stage: , The process simulates the specific recognition process of the enzyme on the substrate, so that the characterization of the enzyme is specifically adjusted according to the chemical property of the specific substrate.
5. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model according to claim 1, wherein in step four the hybrid expert network is a geometric sense hybrid expert module; By pooling active site residues and combining with the conditional features of the substrate recognition stage Generating geometric routing vectors for guiding subsequent expert selection Specifically, define Index sets for residues belonging to active site pockets in the enzyme sequence; integrating the identification signal with the geometric signal by a joint pooling strategy: , Wherein, the The feature is represented by a mean-pooling operation, Representing vector concatenation, routing vector The method is used for reflecting the current substrate binding state and the local pocket geometric topology at the same time, and provides decision basis for subsequent conformation adaptation.
6. The method for predicting the dynamic parameters of the multi-modal enzyme adapted to the protein language model according to claim 1, wherein in the step five, the enzyme-substrate distribution alignment module calculates the distribution distance of the constraint of the maximum mean difference index by a gaussian kernel function, and constrains the features in the same semantic manifold, specifically, defines global feature representations of different stages: , Wherein the method comprises the steps of Respectively corresponding to the initial sequence, MRCA output and G-MoE output by using Gaussian kernel function Calculating the maximum mean difference: ; By minimizing this distribution alignment loss, pre-trained protein language model biochemical semantic catastrophic forgetting is avoided.
7. The method for predicting the dynamic parameters of the multi-modal enzyme adapted to the protein language model according to claim 1, wherein the regression prediction head in the step six is a heteroscedastic gaussian regression prediction head; deep features to adapt the conformation to the phase Inputting the heteroscedastic Gaussian regression prediction head, and outputting a corresponding enzyme kinetic parameter prediction result and uncertainty estimation; specifically, the prediction head output average And logarithmic variance : , Wherein the predicted value of the target kinetic parameter is , Used to construct the heteroscedastic gaussian negative log-likelihood loss.
8. The method for predicting the dynamic parameters of the multi-modal enzyme adapted to the protein language model according to claim 1, wherein in the seventh step, the regression task loss is a heteroscedastic gaussian negative log-likelihood loss, the distribution alignment loss is an MMD loss, and the three are weighted by preset weights to form a comprehensive loss function; specific total loss function Lost by regression task Expert load balancing penalty Loss of distributed alignment The composition is as follows: , 。
9. The method for predicting multimodal enzyme kinetic parameters adapted to a protein language model according to claim 1, wherein the training process in step eight is used for maintaining the equilibrium of the expert network utilization and the stability of the characteristic spatial distribution; and the model after training realizes the prediction of parameters to be measured by processing the unknown enzyme related data.

Description

Multi-modal enzyme dynamics parameter prediction method adapting to protein language model Technical Field The invention relates to the technical field of artificial intelligence and bioinformatics intersection, in particular to a multi-modal enzyme dynamics parameter prediction method adapting to a protein language model. Background The enzyme is used as a core functional element in a biocatalysis process, and kinetic parameters (such as a conversion number kcat, a Mie constant Km, an inhibition constant Ki and the like) of the enzyme directly quantify the catalytic efficiency and the substrate affinity, so that the enzyme is key basic data for screening synthetic biological elements, optimizing metabolic pathways and developing targeted drugs. The method for measuring the enzyme kinetic parameters by the traditional experiment is time-consuming, labor-consuming, high in cost and difficult to meet the application requirements of high-throughput screening. Along with the deep penetration of artificial intelligence technology in the field of computational biology, an enzyme kinetic parameter prediction model based on deep learning becomes an important research direction for replacing the traditional experimental method, and the parameter acquisition efficiency is remarkably improved. In recent years, an enzyme kinetic parameter prediction method is gradually developed from an early shallow machine learning model to a deep learning framework fused with multi-modal information through multiple technological iterations, three major technological paths are mainly formed, namely, firstly, the method is based on a shallow fusion framework of sequence and molecular structure, the preliminary combination of multi-modal information is realized, but the feature interaction depth is insufficient, secondly, the deep evolution and biochemical semantic features of an enzyme sequence are obtained through a feature freezing or fine tuning mode based on a prediction framework of a pre-trained Protein Language Model (PLM), the prediction precision is greatly improved, the method becomes a current mainstream technical scheme, thirdly, a geometric enhancement framework of three-dimensional structure information is introduced, three-dimensional geometric topological features of enzyme active sites are extracted through models such as an isograph neural network, the method tries to make up for the defect of one-dimensional sequence representation on spatial structure information, and modeling capability of the model on an enzymatic reaction mechanism is expected to be further improved. Although the prior art has advanced stepwise, the technology has technical limitations which are difficult to break through in the aspects of simulating a real enzymatic reaction process, excavating geometric structure value, guaranteeing multi-mode fusion stability and the like, and seriously restricts the prediction precision and generalization capability of a model, and the technology is characterized in that the modeling of an enzymatic reaction staged dynamic mechanism is missing, the real enzymatic reaction needs to undergo a dynamic process of substrate specificity recognition-active site conformation adaptation-catalytic reaction occurrence, the existing multi-mode prediction model generally adopts a static fusion strategy, and the differential modeling of two key stages of substrate recognition and conformation adaptation is not carried out no matter the simple feature splicing or the single cross attention interaction. The static modeling thought cannot capture the dynamic conformational change of the enzyme active site after the substrate is combined, so that the model is difficult to accurately describe the specific interaction mechanism of the enzyme and the substrate, and the predicted result has deviation from the actual enzymatic reaction rule. Active site geometric heterogeneity is underutilized in that the active site of an enzyme has significant structural diversity and conformational flexibility, and when different substrates are combined, specific local conformational adjustment of the active site pocket occurs to adapt to the substrate structure. The existing method for introducing three-dimensional structure information mostly adopts a single shared network module to process the geometric characteristics of all active sites, and a differential characteristic extraction mechanism cannot be designed aiming at pocket structures with different geometric topologies. Especially, the self-adaptive modeling strategy capable of dynamically matching the conformational change of the pocket is lacking, the regulation and control value of geometric structure information on enzymatic reaction cannot be fully mined, the expected performance improvement effect cannot be achieved due to the introduction of three-dimensional structure information, and the generalization of the model is reduced due to the feature redundancy even in partial scen