Enzyme isoelectric point and optimal condition prediction method based on multi-channel sequence structure fusion
By using a multi-channel sequence structure fusion model, the problems of limited coverage and insufficient robustness in enzyme parameter prediction in existing technologies are solved, and efficient prediction of enzyme isoelectric point and optimal conditions is achieved, which is applicable to enzyme resource screening and process optimization.
Patent Information
- Application Number
- CN202511766240.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-02-24
AI Technical Summary
Existing technologies mostly focus on a single objective, have limited coverage, lack effective integration of structural/surface information with traditional physicochemical characteristics, and lack robust design in the case of modality absence, making it difficult to support rapid prediction of high-throughput enzyme parameters and process optimization.
A multi-channel sequence structure fusion model was constructed. Feature vectors of multi-source data were extracted through a protein language model. Combined with traditional sequence statistics, ESM2 pre-trained semantics, and ProtBert surface residue level channels, the enzyme isoelectric point and optimal conditions were predicted in a unified manner. The model parameters were optimized by cross-validation on the training set.
It enables simultaneous prediction of enzyme isoelectric point, optimal pH, and optimal temperature, improving the model's robustness and high-throughput processing capabilities, reducing experimental costs and time, and is suitable for enzyme resource screening, buffer system setting, and process optimization.
Smart Images
Figure CN121565242A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of bioinformatics and computational enzyme engineering technology, and particularly relates to a method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion. Background Technology
[0002] Prior to in vitro experiments, it is necessary to calculate, predict, and pre-screen key experimental parameters of enzymes. These key parameters include: isoelectric point (pI), the pH at which the overall net charge of the protein is zero; optimal pH (pHopt), the acidity or alkalinity at which the enzyme's catalytic activity reaches its peak; and optimal temperature (Topt), the temperature at which the activity is highest under given buffer system and substrate conditions. These parameters directly relate to buffer system formulation design, fermentation / reaction condition setting, mutant selection, and determination of the large-scale process window. Accurate and rapid acquisition of these parameters can significantly reduce the cost and cycle time of experimental iterations.
[0003] In existing technologies, the aforementioned experimental parameters mainly rely on wet-method experiments: for example, pI is often estimated through isoelectric focusing or electrophoresis, while the optimal pH / temperature is obtained through activity scanning curves or steady-state enzyme kinetic experiments. Although these methods have high accuracy, they are costly, time-consuming, and significantly affected by the heterogeneity of experimental conditions (such as differences in buffer salt concentration, ionic strength, and cofactor and substrate concentrations), making it difficult to support high-throughput pre-evaluation of massive amounts of enzyme sequences. Therefore, using computational methods to provide reliable parameter ranges or recommended values before experiments has become an important trend and a practical need in this field.
[0004] In predicting enzyme isoelectric point and optimal conditions, pIR in the pI direction extracts amino acid composition and physicochemical descriptors, such as hydrophobicity, charge, dipeptide frequency, and CTD, as input from the sequence. After standardization and feature selection, it is modeled using SVM regression. IPC1.0 transforms pI calculation into a global pKa scale fitting problem within the Henderson–Hasselbalch framework, simultaneously optimizing the pKa of the terminal and side chains to match the sequence charge neutralization point. IPC2.0 integrates this physicochemical constraint with SVM / deep regression to achieve integrated learning of "theoretical priors + data-driven". In the pHopt direction, CatOpt uses ESM-2 to encode sequences into dense representations, which are then fed into a multi-scale CNN with multiple receptive fields / dilated convolutions, and combined with attention and residual-dense modules for deep feature fusion. Venus-DREAM adopts "retrieval-enhanced meta-learning": first, the ESM-2 representation is used to retrieve the Top-k nearest neighbors according to cosine similarity to form a support set, and then the encoder is frozen on the EpHod backbone (ESM-1v encoder + lightweight regression head), and only the prediction head is subjected to Reptile few-sample adaptation. During inference, a small number of updates are also made on the support set first. Seq2pHopt uses ESM-2 and ProGen2 dual encoders for parallel extraction, and the global representation is stacked into 4 residual-dense blocks after multi-head attention alignment. In the Topt direction, DeepET first performs transfer / self-supervised pre-training on a large-scale sequence-phenotype model (using the optimal growth temperature of the organism as a proxy for the optimal temperature of the enzyme) to obtain a general representation related to temperature. Then, it transfers this representation to enzyme Topt prediction and adds a shallow regression head. Seq2Topt continues the idea of "pre-trained representation + residual-dense block", using large model embeddings such as ESM-2 / ProGen2 as a base, and regresses temperature through deep nonlinear modules.
[0005] Overall, existing technologies mostly focus on a single objective, and publicly available datasets have limited coverage and inconsistent definitions. Most methods still rely on single-modal modeling of sequences, lacking effective integration of multi-source data such as structural / surface information and traditional physicochemical characteristics, and also lacking robust designs in the case of missing modalities. Although progress has been made in various tasks, there is still insufficient support for uncertainty characterization and engineering-based high-throughput deployment, and overall performance still has room for improvement.
[0006] Therefore, this invention aims to construct a multi-channel fusion model that can simultaneously predict pI, pHopt, and Topt within a unified framework. Using amino acid sequences as basic inputs, and combining three-dimensional structure / surface information with traditional physicochemical characteristics, it achieves automatic prediction of the isoelectric point (pI), optimum pH (pHopt), and optimum temperature (Topt) of enzymes. It also has efficient reasoning capabilities for long sequences and large batches of samples, and can be applied to enzyme resource screening, buffer system and fermentation / reaction condition setting, mutant selection, and process optimization. Summary of the Invention
[0007] The purpose of this invention is to provide a method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion, in order to solve the problems that existing technologies mostly focus on a single target, have limited coverage and inconsistent standards of publicly available datasets, are still limited to single-modal sequence modeling, lack effective fusion of multi-source data such as structural / surface information and traditional physicochemical characteristics, and lack the technical problem of robust design in the case of modality missing.
[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion includes the following steps: S1: Obtain multi-source data including the target variable from enzymological records, align and standardize the multi-source data with a unified protein identifier, and divide it into training set and test set according to a specified ratio; S2: Multi-channel feature extraction is performed on the training set using the constructed protein language model to obtain the feature vectors for each channel. S3: The feature vectors of each channel are subjected to lightweight linear mapping and nonlinear activation to obtain a medium-dimensional representation; S4: The medium-dimensional representations are concatenated in a fixed order to form a fusion vector. The key parameters of the enzyme, including the enzyme's isoelectric point and optimal conditions, are output through a two-layer perceptron. The optimal parameter combination is obtained by five-fold cross-validation on the training set. S5: Input the test set into the trained protein language model and output the predicted value. Set multiple evaluation indicators to evaluate the predicted value and determine whether each evaluation result is within the preset fluctuation range. If yes, output the predicted result; otherwise, retrain the model.
[0009] Preferably, the specific process of obtaining multi-source data including the target variable from enzymatic records in step S1, and aligning and standardizing the multi-source data with a unified protein identifier is as follows: S11: First, obtain enzymatic records from BRENDA in batches, including the target variables such as isoelectric point pI, optimal pH pHopt, and optimal temperature Topt. S12: Using UniProtKB Accession as the main bond, extract the corresponding amino acid primary sequence from UniProt; S13: Structural information is downloaded from the AlphaFold Protein Structure Database according to priority, and missing entries are filled in by ESMFold prediction; For multiple observations of the same protein, an arithmetic mean was used to generate a unique label; the temperature was uniformly set to degrees Celsius, and pH and pI were dimensionless. S14: Perform data cleaning, remove null values, missing values and significant anomalies, and deduplicate data by Accession.
[0010] Preferably, in step S2, the protein language model is set up with three parallel channels to process the isoelectric point pI, the optimum pH pHopt, and the optimum temperature Topt separately. The first channel is the traditional sequence statistics and physicochemical channel, which performs low-order prior operations on mining length, composition and hydrophobicity. The second channel is the ESM2 pre-trained semantic channel, which uses a protein language model to extract a globally distributed representation. The third channel is the ProtBert surface residue level channel. First, FreeSASA determines the surface residues, and then ProtBert's residue-by-residue embedding is used to perform attention convergence on the surface subset.
[0011] Preferably, the specific feature extraction process for the first channel is as follows: Let the protein sequence be , For the 20 standard amino acid alphabets, illegal characters are mapped as follows: ; S21: Extract length-related features and amino acid composition vector ,in in, The sequence length; For the first The amino acid symbol of each residue; This represents a set of 20 standard amino acids; Indicates unknown / non-standard residues; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. S22: According to a predefined amino acid set The proportions of the 12 categories—acidic, basic, charged, polar / uncharged, nonpolar, hydrophobic, aromatic, aliphatic, tiny, small, pro, and gly—were statistically analyzed. in, For the predefined first A set of amino acid groups, with the remaining symbols as above, each... This indicates the proportion of the group in the entire sequence; S23: Kyte–Doolittle hydrophobic sequence The statistics include mean ; Standard deviation ; Minimum value ; Maximum value ; in, Kyte–Doolittle hydrophobic scale; and These are the mean and standard deviation, respectively. These are the minimum / maximum scale values; S24: Molecular weight approximation based on residue mass tables ; The simplified CTD (Composition–Transition–Distribution) is as follows: Where X = (x_1,…,x_L) is an amino acid sequence of length L; i is a residue index; |{ } | represents the number of elements in the set; H( ) represents the Kyte–Doolittle hydrophobicity scale; t is the selected hydrophobicity threshold used to distinguish between three categories: hydrophobic (H>t), hydrophilic (H<-t), and neutral (-t≤H≤t); ctd_hydrophob, ctd_hydrophil, and ctd_hydroneut represent the proportions of the three types of residues in the sequence, and the sum of the three is always 1.
[0012] Preferably, the specific feature extraction process for the second channel is as follows: Let the input sequence be ESM2 encoding yields the final hidden state: Take the CLS position as the sequence-level feature vector: ; in, For the last layer of ESM2 The hidden vector at each position, The CLS position vector is directly used as the second channel output.
[0013] Preferably, the specific feature extraction process for the third channel is as follows: S25: Let the first The solvent-accessible area of each residue is Its maximum surface area constant is The relative solvent accessibility is defined as: ; in, For the first Solvent-accessible area of each residue This represents the maximum accessible area of the residue in the standard reference configuration. This refers to relative solvent accessibility; in, Surface threshold; It is a set of surface residue indexes; For length is A binary mask; S26: Input the same amino acid sequence into ProtBert to obtain a residue-by-residue hidden representation: in, For the first Context-dependent vectors for each amino acid; right Calculate the weights of the vectors within and aggregate them: in, For the first Scoring of each residue; These are learnable parameters; Used to calculate weights, without changing The dimension; S27: After masking the non-surface locations, perform masked softmax on the scores to obtain the attention weights: in, For the first The normalized weights of residues; if but ; S28: Obtain the surface attention pooling vector: in: With the same dimension as the single residue embedding (1024 dimensions), it serves as the final output of this channel and directly enters the subsequent fusion layer; It can also be used as an interpretability weight to locate key surface sites.
[0014] Preferably, the specific process of performing lightweight linear mapping and nonlinear activation on the feature vectors of each channel in step S3 is as follows: Let the sequence statistical characteristics be denoted as ESM2 features are Surface pooling characteristics are (1024 dimensions), then: ; ; in, These are learnable projection parameters.
[0015] Preferably, the specific process in step S4 is as follows: The three aligned representations are concatenated in a fixed order to form a fused vector, which is then used for regression prediction via a two-layer perceptron. ; ; ; in, For layer normalization; As a learnable parameter, Dropout is set to 0.3; This is a single-task output at a standardized scale.
[0016] Preferably, the specific process of setting multiple evaluation indicators and evaluating the predicted value in step S5 is as follows: Assume there are a total of For each sample, the actual value and the predicted value are respectively... and , record the mean , ; S51: Calculate MSE: ; S52: Calculate RMSE: ; S53: Calculation : ; S54: Calculation : ; S55: Calculate NRMSE: .
[0017] The beneficial effects of this invention include: The present invention provides a method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion. 1. To address the technical problems of existing methods relying on small-scale datasets from many years ago, resulting in incomplete coverage of enzyme families and species; inconsistent data definitions; and the fact that most methods only use single-modality sequences and lack structural information support, this invention systematically integrates enzyme / reaction information from BRENDA and SABIO-RK, and uses UniProt to obtain standardized sequences; it batch downloads corresponding 3D structures from AlphaFoldDB, and uses ESMFold prediction to fill in missing sequences; at the same time, it completes definition alignment, unit standardization, and redundancy removal to construct a more comprehensive and reusable training / evaluation dataset, which improves model learnability and facilitates reproduction and expansion.
[0018] 2. To address the technical problem that existing methods are mostly single-objective and single-modal modeling of sequences, and that the structure / surface and physicochemical features are not effectively integrated, making it difficult to make full use of complementary information, this invention proposes a method and system for predicting enzyme isoelectric point and optimal conditions by multi-channel information fusion: under the premise of sequence as the main focus, the method adaptively integrates structure / surface and physicochemical channels to simultaneously predict the isoelectric point (pI), optimal pH (pHopt), and optimal temperature (Topt) within a unified framework; the method is robust to modality loss and is compatible with long sequences and batch high-throughput inference.
[0019] 3. Common distribution shifts between training and application scenarios lead to insufficient prediction stability; the lack of robust design for missing or low-quality structures makes it difficult to support large-scale screening and rapid setting of process parameters, hindering effective reduction of experimental costs and timelines. This invention improves accuracy and stability (increased R², reduced RMSE / MAE), and automatically completes missing or low-quality structural modes through built-in structure prediction and surface feature inference modules, maintaining the integrity of multi-channel information and significantly enhancing robustness. Furthermore, in terms of engineering, it enables high-throughput long-sequence processing with limited GPU memory, and combined with internal validation and early stopping to save the best model, ensures consistency and reproducibility. Attached Figure Description
[0020] Figure 1 This is a flowchart illustrating the method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion according to the present invention.
[0021] Figure 2 This is a schematic diagram illustrating the principle of the present invention for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion.
[0022] Figure 3 The figure shows the results of five-fold cross-validation and test set for the prediction of the three enzyme parameters in this invention. Detailed Implementation
[0023] The following is in conjunction with the appendix Figures 1-3 The present invention will be further described in detail below: Example 1 See appendix Figure 1 and Figure 2 As shown, the method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion includes the following steps: S1: Obtain multi-source data including target variables from enzymological records. The target variables include isoelectric point (pI), optimum pH (pHopt), and optimum temperature (Topt). Align and standardize the multi-source data with a unified protein identifier, and divide it into training and test sets according to a specified ratio of 9:1.
[0024] S2: Multi-channel feature extraction is performed on the training set using a constructed protein language model, yielding feature vectors for each channel. The protein language model takes the sequence and structure of the same protein as input and constructs three parallel channels: the first is a traditional sequence statistics and physicochemical channel, mining low-order priors such as length, composition, and hydrophobicity; the second is an ESM2 pre-trained semantic channel, extracting a globally distributed representation using a large-scale protein language model; and the third is a ProtBert surface residue-level channel, first using FreeSASA to determine surface residues, then using ProtBert's residue-by-residue embedding to perform attention convergence on surface subsets, thereby strengthening surface information closely related to environmental interactions. Each of the three channels generates a fixed-dimensional vector, which is then fused into the regression head after a subsequent uniform linear transformation to output the final predicted value.
[0025] S3: The feature vectors of each channel are subjected to lightweight linear mapping and nonlinear activation to obtain a medium-dimensional representation with consistent magnitude and more stable numerical distribution.
[0026] S4: The medium-dimensional representations are concatenated in a fixed order to form a fusion vector. The key parameters of the enzyme, including the enzyme's isoelectric point and optimal conditions, are output through a two-layer perceptron. The optimal parameter combination is obtained by five-fold cross-validation on the training set.
[0027] S5: Input the test set into the trained protein language model and output the predicted value. Set multiple evaluation indicators to evaluate the predicted value and determine whether each evaluation result is within the preset fluctuation range. If yes, output the predicted result; otherwise, retrain the model.
[0028] The specific process of obtaining multi-source data including the target variable from enzymatic records in step S1, and aligning and standardizing the multi-source data with a unified protein identifier is as follows: S11: First, obtain enzymatic records containing target variables such as pI, pHopt, and Topt from BRENDA in batches; S12: Using UniProtKB Accession as the main bond, extract the corresponding amino acid primary sequence from UniProt; S13: Structural information is downloaded from the AlphaFold Protein Structure Database according to priority, and missing entries are filled in by ESMFold prediction; For multiple observations of the same protein, an arithmetic mean was used to generate a unique label; the temperature was uniformly set to degrees Celsius, and pH and pI were dimensionless. S14: Perform data cleaning, remove null values, missing values and significant anomalies, and deduplicate data by Accession.
[0029] Example 2 Based on Example 1, in step S2, the protein language model sets up three parallel channels to process the isoelectric point pI, optimum pH pHopt, and optimum temperature Topt separately. Each of the three channels generates a fixed-dimensional vector, which is then fused into the regression head after a uniform linear transformation to output the final predicted value. The first channel is a traditional sequence statistics and physicochemical channel, which performs low-order prior operations to mine length, composition, and hydrophobicity; the second channel is an ESM2 pre-trained semantic channel, which uses a large-scale protein language model to extract a globally distributed representation; the third channel is a ProtBert surface residue hierarchy channel, which first uses FreeSASA to determine surface residues, and then uses ProtBert's residue-by-residue embedding to perform attention convergence on the surface subset, thereby strengthening the surface information closely related to environmental interactions.
[0030] The specific feature extraction process for the first channel is as follows: Let the protein sequence be , For the 20 standard amino acid alphabets, illegal characters are mapped as follows: ; S21: Extract length-related features and amino acid composition vector ,in in, The sequence length; For the first The amino acid symbol of each residue; This represents a set of 20 standard amino acids; Representing unknown / non-standard residues (uniformly mapped to) ); This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. S22: According to a predefined amino acid set The proportions of the 12 categories—acidic, basic, charged, polar / uncharged, nonpolar, hydrophobic, aromatic, aliphatic, tiny, small, pro, and gly—were statistically analyzed. in, For the predefined first A set of amino acid groups, sequentially covering acidic, basic, charged, polar / uncharged, nonpolar, hydrophobic, aromatic, aliphatic, tiny, small, pro, and gly, with the remaining symbols being the same as above. This indicates the proportion of the group in the entire sequence; S23: Kyte–Doolittle hydrophobic sequence The statistics include mean ; Standard deviation ; Minimum value ; Maximum value ; in, Kyte–Doolittle hydrophobic scale; and These are the mean and standard deviation, respectively. These are the minimum / maximum scale values; S24: Molecular weight approximation based on residue mass tables .
[0031] The simplified CTD (Composition–Transition–Distribution) is as follows: Where X = (x_1,…,x_L) is an amino acid sequence of length L; i is a residue index; |{ } | represents the number of elements in the set; H( `t` represents the Kyte–Doolittle hydrophobicity scale; `t` is the selected hydrophobicity threshold used to distinguish between hydrophobic (H>t), hydrophilic (H<-t), and neutral (-t≤H≤t); `ctd_hydrophob`, `ctd_hydrophil`, and `ctd_hydroneut` represent the proportions of the three types of residues in the sequence, and their sum is always 1. This channel directly provides a 45-dimensional real-valued vector as subsequent input.
[0032] The specific feature extraction process for the second channel is as follows: The ESM2 model with model number esm2_t30_150M_UR50D is used to encode the sequence, and the global representation of the last layer is directly taken as the feature without any dimensional changes. Let the input sequence be ESM2 encoding yields the final hidden state: Take the CLS position as the sequence-level feature vector: ; in, For the last layer of ESM2 The hidden vector at each position, The CLS position vector is directly used as the output of this channel.
[0033] The specific feature extraction process for the third channel is as follows: To highlight the environmental interaction information carried by the structural surface, the relative solvent accessible area (rSASA) of the residues is calculated for each three-dimensional structure. S25: Let the first The solvent-accessible area of each residue is Its maximum surface area constant is , The relative solvent accessibility, which can be given by a free-sphere probe model or standard table, is defined as follows: ; in, For the first Solvent-accessible area of each residue This represents the maximum accessible area of the residue in the standard reference configuration; both were calculated using FreeSASA. Relative accessibility (percentage); in, Surface threshold (example taken) ); It is a set of surface residue indexes; For length is A binary mask; In extreme cases ,make To avoid an empty set; S26: Input the same amino acid sequence into ProtBert to obtain a residue-by-residue hidden representation: in, For the first The context-related vector of each amino acid has a constant dimension of 1024. This step does not change the dimension or perform projection.
[0034] To focus on discriminative information for surface regions, only for Calculate the weights of the vectors within and aggregate them: in, For the first Scoring of each residue; These are learnable parameters; Used only for calculating weights, does not change The dimension; S27: After masking the non-surface locations (assigning large negative numbers in the implementation), perform masked softmax on the scores to obtain the attention weights: in, For the first The normalized weights of residues; if but ; S28: Obtain the surface attention pooling vector: in: With the same dimension as the single residue embedding (1024 dimensions), it serves as the final output of this channel and directly enters the subsequent fusion layer; It can also be used as an interpretability weight to locate key surface sites.
[0035] This design uses a geometric mask generated by rSASA to constrain attention, weighting only the surface region. This can suppress the noise contribution of buried residues to the prediction and highlight the site features that directly interact with the solvent and ionic environment. At the same time, it maintains 1024 dimensions from input to output, which is convenient for alignment with other channels and engineering deployment.
[0036] Example 3 Based on Example 1 or Example 2, the specific process of performing lightweight linear mapping and nonlinear activation on the feature vectors of each channel in step S3 is as follows: Let the sequence statistical characteristics be denoted as ESM2 features are (CLS from the previous section, 640-dimensional), surface pooling characteristics are: (1024 dimensions), then ; ; in, The projected parameters are learnable; the typical setting projects the three paths to 256 dimensions, and adds LayerNorm before projection and Dropout (0.3) after ReLU to improve generalization.
[0037] The specific process in step S4 is as follows: The three aligned representations are concatenated in a fixed order to form a fused vector, which is then used for regression prediction via a two-layer perceptron. in, For layer normalization; These are learnable parameters; the hidden dimension is typically 512, and Dropout is 0.3. This is a single-task output at a standardized scale.
[0038] Step S5 involves setting multiple evaluation indicators to evaluate the predicted values. The specific process is as follows: Assume there are a total of For each sample, the actual value and the predicted value are respectively... and , record the mean , ; S51: Calculate MSE (mean squared error, the smaller the better): ; S52: Calculate RMSE (Root Mean Square Error, ↓): ; S53: Calculation (Determination coefficient, ↑): S54: Calculation (Pearson correlation coefficient, ↑): S55: Calculate NRMSE (Normalized RMSE, ↓): .
[0039] The model of this invention shows that the key indicators of five-fold cross-validation and independent testing are basically consistent across the three regression tasks (pI, pHopt, and Topt), demonstrating stable generalization ability and transferability: on the independent test set, the R^2 and r of pI are approximately 0.661 / 0.813, pHopt is approximately 0.444 / 0.666, and Topt is approximately 0.517 / 0.719; the error indicators (RMSE / NRMSE) are all within the reasonable fluctuation range of the cross-validation mean, and no obvious overfitting is observed.
[0040] For the isoelectric point (pI): The mean values of the five-fold cross-validation were: MSE=0.690, RMSE=0.830, R^2=0.659, r=0.812, NRMSE=0.584 (see Table 1 and...). Figure 3 (AB). On the independent test set, the MSE was 0.736, RMSE was 0.858, R² was 0.661, r was 0.813, and NRMSE was 0.582 (see Table 1). Figure 3 B. The test results are highly consistent with the cross-validation mean, indicating that the model fits pI stably and has a strong linear correlation.
[0041] Table 1. Results of pI five-fold cross-validation and test set For the optimal pH (pHopt): See Table 2 and Figure 3 The overall performance of the CD five-fold cross-validation was: MSE=0.744, RMSE=0.862, R^2=0.421, r=0.649; NRMSE=0.761. See Table 2 and... Figure 3 D. On the independent test set, the results showed MSE=0.740, RMSE=0.860, R^2=0.444, r=0.666, and NRMSE=0.746. The cross-validation and test performance were close, indicating that the method generalizes stably on pHopt.
[0042] Table 2. Results of pHopt five-fold cross-validation and test set. For the optimal temperature (Topt): See Table 3 and Figure 3 The mean values for EF and the five-fold cross-validation are: MSE = 119.551, RMSE = 10.934, R² = 0.543, r = 0.737; NRMSE = 0.676. See Table 3 and... Figure 3F, the results on the independent test set are: MSE=121.300, RMSE=11.014, R^2=0.517, r=0.719, NRMSE=0.695. The test and cross-validation mean values are comparable, indicating that it also has good robustness and transferability on the Topt task.
[0043] Table 3. Results of Topt five-fold cross-validation and test set. In summary, the enzyme isoelectric point and optimal condition prediction method based on multi-channel sequence structure fusion provided by this invention proposes a multi-channel sequence-structure surface fusion prediction method and system for enzyme physicochemical optimal conditions (pI / pHopt / Topt). Using UniProtKB Accession as the unified primary key, the target tags of BRANDA are aligned with UniProt sequences and AlphaFold DB structures. When a structure is not hit, it is automatically filled in by ESMFold. Three complementary feature channels are constructed in parallel at the modeling end: a statistical / physicochemical prior channel based on the first-level sequence (45 dimensions, including length, amino acid composition, 12 groups of physicochemical groups, KD statistics, molecular weight and simplified CTD three dimensions); a CLS global semantic channel (640 dimensions) extracted by ESM2 (facebook / esm2_t30_150M_UR50D); and a surface residue hierarchical channel that performs masking attention / mean pooling only on the surface residue set using ProtBert (Rostlab / prot_bert) as the backbone and combined with rSASA threshold calculated by FreeSASA (1024-dimensional embedding per residue, 1024-dimensional output). The three-channel features are concatenated after lightweight projection / normalization, and the target value is output by a regression head consisting of an LN and two MLP layers. During training, the Z-score parameters are recorded only for predicting the inverse transform, while the loss is calculated on the true scale (MSE). This process ensures practical feasibility while significantly improving generalization and robustness to cross-family / new species samples by utilizing "geometrically constrained surface features + pre-trained semantics," and can output surface attention weights for site interpretation and experimental design.
Claims
1. A method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion, characterized in that, Includes the following steps: S1: Obtain multi-source data including the target variable from enzymological records, align and standardize the multi-source data with a unified protein identifier, and divide it into training set and test set according to a specified ratio; S2: Multi-channel feature extraction is performed on the training set using the constructed protein language model to obtain the feature vectors for each channel. S3: The feature vectors of each channel are linearly mapped and nonlinearly activated to obtain a medium-dimensional representation; S4: The medium-dimensional representations are concatenated in a fixed order to form a fusion vector. The key parameters of the enzyme, including the enzyme's isoelectric point and optimal conditions, are output through a two-layer perceptron. The optimal parameter combination is obtained by five-fold cross-validation on the training set. S5: Input the test set into the trained protein language model and output the predicted value. Set multiple evaluation indicators to evaluate the predicted value and determine whether each evaluation result is within the preset fluctuation range. If yes, output the predicted result; otherwise, retrain the model.
2. The method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion according to claim 1, characterized in that, The specific process of obtaining multi-source data including the target variable from enzymatic records in step S1, and aligning and standardizing the multi-source data with a unified protein identifier is as follows: S11: First, obtain enzymatic records from BRENDA in batches, including the target variables such as isoelectric point pI, optimal pH pHopt, and optimal temperature Topt. S12: Using UniProtKB Accession as the main bond, extract the corresponding amino acid primary sequence from UniProt; S13: Structural information is downloaded from the AlphaFold Protein Structure Database according to priority, and missing entries are filled in by ESMFold prediction; For multiple observations of the same protein, an arithmetic mean is used to generate a unique label; Temperature is uniformly measured in degrees Celsius, and pH and pI are dimensionless. S14: Perform data cleaning, remove null values, missing values and significant anomalies, and deduplicate data by Accession.
3. The method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion according to claim 2, characterized in that, In step S2, the protein language model is set up with three parallel channels to process the isoelectric point pI, the optimum pH pHopt, and the optimum temperature Topt respectively. The first channel is the traditional sequence statistics and physicochemical channel, which performs low-order prior operations on mining length, composition and hydrophobicity. The second channel is the ESM2 pre-trained semantic channel, which uses a protein language model to extract a globally distributed representation. The third channel is the ProtBert surface residue level channel. First, FreeSASA determines the surface residues, and then ProtBert's residue-by-residue embedding is used to perform attention convergence on the surface subset.
4. The method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion according to claim 3, characterized in that, The specific feature extraction process for the first channel is as follows: Let the protein sequence be , For the 20 standard amino acid alphabets, illegal characters are mapped as follows: ; S21: Extract length-related features and amino acid composition vector ,in in, The sequence length; For the first The amino acid symbol of each residue; This represents a set of 20 standard amino acids; Indicates unknown / non-standard residues; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. S22: According to a predefined amino acid set The proportions of the 12 categories—acidic, basic, charged, polar / uncharged, nonpolar, hydrophobic, aromatic, aliphatic, tiny, small, pro, and gly—were statistically analyzed. in, For the predefined first A set of amino acid groups, with the remaining symbols as above, each... This indicates the proportion of the group in the entire sequence; S23: Kyte–Doolittle hydrophobic sequence The statistics include mean ; Standard deviation ; Minimum value ; Maximum value ; in, Kyte–Doolittle hydrophobic scale; and These are the mean and standard deviation, respectively. These are the minimum / maximum scale values; S24: Molecular weight approximation based on residue mass tables ; The simplified CTD is as follows: Where X = (x_1,…,x_L) is an amino acid sequence of length L; i is a residue index; |{ } | represents the number of elements in the set; H( ) represents the Kyte–Doolittle hydrophobicity scale; t is the selected hydrophobicity threshold used to distinguish between three categories: hydrophobic (H>t), hydrophilic (H<-t), and neutral (-t≤H≤t); ctd_hydrophob, ctd_hydrophil, and ctd_hydroneut represent the proportions of the three types of residues in the sequence, and the sum of the three is always 1.
5. The method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion according to claim 3, characterized in that, The specific feature extraction process for the second channel is as follows: Let the input sequence be ESM2 encoding yields the final hidden state: Take the CLS position as the sequence-level feature vector: ； in, For the last layer of ESM2 The hidden vector at each position, The CLS position vector is directly used as the second channel output.
6. The method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion according to claim 3, characterized in that, The specific feature extraction process for the third channel is as follows: S25: Let the first The solvent-accessible area of each residue is Its maximum surface area constant is The relative solvent accessibility is defined as: ； in, For the first Solvent-accessible area of each residue This represents the maximum accessible area of the residue in the standard reference configuration. This refers to relative solvent accessibility; in, Surface threshold; It is a set of surface residue indexes; For length is A binary mask; S26: Input the same amino acid sequence into ProtBert to obtain a residue-by-residue hidden representation: in, For the first Context-dependent vectors for each amino acid; right Calculate the weights of the vectors within and aggregate them: in, For the first Scoring of each residue; These are learnable parameters; Used to calculate weights, without changing The dimension; S27: After masking the non-surface locations, perform masked softmax on the scores to obtain the attention weights: in, For the first The normalized weights of residues; if but ; S28: Obtain the surface attention pooling vector: in: With the same dimension as the single residue embedding (1024 dimensions), it serves as the final output of this channel and directly enters the subsequent fusion layer; It can also be used as an interpretability weight to locate key surface sites.
7. The method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion according to claim 1, characterized in that, The specific process of performing lightweight linear mapping and nonlinear activation on the feature vectors of each channel in step S3 is as follows: Let the sequence statistical characteristics be denoted as ESM2 features are Surface pooling characteristics are (1024 dimensions), then: ； ； in, These are learnable projection parameters.
8. The method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion according to claim 1, characterized in that, The specific process in step S4 is as follows: The three aligned representations are concatenated in a fixed order to form a fused vector, which is then used for regression prediction via a two-layer perceptron. ； ； ； in, For layer normalization; As a learnable parameter, Dropout is set to 0.3; This is a single-task output at a standardized scale.
9. The method for predicting enzyme isoelectric point and optimal conditions based on multi-channel sequence structure fusion according to claim 1, characterized in that, Step S5 involves setting multiple evaluation indicators to evaluate the predicted values. The specific process is as follows: Assume there are a total of For each sample, the actual value and the predicted value are respectively... and , record the mean , ; S51: Calculate MSE: ； S52: Calculate RMSE: ； S53: Calculation : ； S54: Calculation : ； S55: Calculate NRMSE: 。