Method for predicting enzyme solubility by using machine learning model
By extracting specific physicochemical features and utilizing a Transformer model, the method enhances enzyme solubility prediction accuracy, addressing the limitations of existing models by capturing long-distance dependencies and improving prediction precision.
Patent Information
- Application Number
- CN202411268197.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-11
- Publication Date
- 2024-12-13
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing enzyme solubility prediction models lack accuracy due to incomplete characterization of complex interactions between amino acids and neglect global features, leading to insufficient solubility predictions that affect protein production and research.
A method involving the extraction of five physicochemical features (AAC, PseAAC, APAAC, QSO, DPC) from enzyme sequences, combined with a Transformer model and fully connected layers to predict solubility, capturing long-distance dependencies and improving prediction accuracy.
The method achieves high-precision enzyme solubility predictions by leveraging multi-dimensional sequence information and the Transformer's attention mechanism, outperforming traditional models in accuracy and efficiency.
Smart Images
Figure CN119132406A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method for predicting enzyme solubility using a machine learning model, and belongs to the technical field of bioinformatics. Background Art
[0002] Protein solubility is of vital importance in the fields of biotechnology and pharmaceuticals. As the basis for heterologous protein expression, protein solubility directly affects the biological activity and function of proteins, and thus determines the success of recombinant protein production. Insufficient solubility will cause proteins to form insoluble aggregates (inclusion bodies) in host cells, which not only affects downstream protein purification and functional analysis, but may also hinder the development of new protein drugs. In addition, protein solubility issues are associated with a variety of human diseases, such as neurodegenerative diseases such as Alzheimer's disease and Parkinson's disease, which further highlights the need to improve the ability to predict protein solubility in protein science research. Although multiple prediction models have been developed to address this problem, the accuracy of existing models is still limited due to the influence of complex biophysical and chemical properties and the intracellular environment. Therefore, the development of more accurate and efficient solubility prediction tools is of great significance to promote protein research and its applications.
[0003] Kodai et al. proposed a general model structure in 2022. They first extracted features related to the enzyme sequence, such as amino acid composition, secondary structure information, specific motifs or patterns in the sequence, and then used the extracted features combined with the LSTM network to predict enzyme solubility. However, the extracted features cannot fully characterize the complex interactions between amino acids in the sequence, and this lack may affect the model's accurate prediction of enzyme solubility. At the same time, the model may over-rely on the local order of the sequence and ignore global changes, and cannot make full use of global features for more robust predictions. For the above reasons, it is necessary to study a method for using a machine learning model to predict enzyme solubility. Summary of the invention
[0004] In order to solve the problems existing in the existing enzyme solubility prediction model, the present invention provides a method for predicting enzyme solubility using a machine learning model, the method comprising the following steps: Step S1: Extract five physicochemical features from enzyme sequences in public datasets, including amino acid composition (AAC), pseudo-amino acid composition (PseAAC), amphiphilic pseudo-amino acid composition (APAAC), quasi-sequence order (QSO), and dipeptide composition (DPC).
[0005] Step S2: The five physical and chemical features extracted above are spliced to form multi-dimensional and rich sequence information.
[0006] Step S3: Input the enzyme sequence features obtained in step S2 into the Transformer model and combine it with the fully connected layer to realize enzyme solubility prediction.
[0007] Optionally, the step S1 includes: The enzyme sequence features are standardized to ensure that different features are on the same scale and improve the training stability of the model.
[0008] Optionally, in addition to using the five physicochemical characteristics, the three-dimensional structural information, topological characteristics, etc. of the enzyme may also be considered.
[0009] Beneficial effects of the present invention: Public data sets are collected, and five physical and chemical characteristics, including amino acid composition (AAC), pseudo amino acid composition (PseAAC), amphiphilic pseudo amino acid composition (APAAC), quasi-sequence order (QSO) and dipeptide composition (DPC), are extracted by designing a specific algorithm, and the features are spliced to form multi-dimensional and rich sequence information. Subsequently, this rich sequence information is input into the Transformer model, and the attention mechanism of the Transformer is used to more accurately capture the long-distance dependencies between amino acids in the sequence. Compared with traditional models, Transformer can process long sequences more efficiently and avoid information loss problems. Finally, by inputting these deeply processed sequence features into the fully connected layer, a high-precision prediction of enzyme solubility is achieved. The present application also provides the application of the above-mentioned machine learning model for predicting enzyme solubility in the fields of chemicals and pharmaceuticals. BRIEF DESCRIPTION OF THE DRAWINGS
[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0011] Figure 1 It is a flowchart of a method for using a machine learning model to predict enzyme solubility provided in one embodiment of the present invention. DETAILED DESCRIPTION
[0012] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0013] PyTorch: is the python version of torch, an open source neural network framework by Facebook, specifically for GPU-accelerated deep neural network (DNN) programming. Torch is a classic tensor library (tensor, a tensor is a digital container in machine learning programs, essentially an array of various dimensions, usually the dimensions of a tensor are called axes, and the number of axes is called order) that operates on multi-dimensional matrix data. It is widely used in machine learning and other math-intensive applications. Unlike the static computational graph of Tensorflow, the computational graph of pytorch is dynamic and can be changed in real time according to computational needs.
[0014] Example: This example provides a method for predicting enzyme solubility using a machine learning model, the method comprising an algorithm for extracting the physicochemical characteristics of an enzyme sequence, a Transformer model, see Figure 1 , the method comprising: Step S1: feature extraction of the input enzyme sequence.
[0015] In model training, the training data set used was extracted from the TargetTrack database, which is a merger of the TargetDB and PepcDB databases. The data set used for model training contains 11,436 proteins (5,718 of which are soluble and 5,718 are insoluble). The training of the model using this data set enables the model to learn the interaction pattern between enzymes and substrates. At the same time, in order to verify these data sets of the model for evaluating the model in different generalization capabilities and predictive performance, the test data set from NESG was used for independent testing of the model. The test data set also underwent the same filtering process and contained 3,100 proteins (1,550 of which were soluble and 1,550 were insoluble).
[0016] Step S1.1: Extracting amino acid composition (AAC) features.
[0017] The frequency of each amino acid in a protein sequence can be used as a feature of the sequence. The frequency of each amino acid in a protein sequence is calculated using the following formula:
[0018] Step S1.2: Extract pseudo-amino acid composition (PseAAC) features.
[0019] The PseAAC feature,descriptor proposed by Chou et al. in 2009 is used to extract the physicochemical information of protein sequences.,PseAAC encodes protein sequences into feature vectors using the following formula:
[0020] Step S1.3: Extracting amphipathic pseudo amino acid composition (APAAC) features.
[0021] APAAC helps the model better understand the solubility and biological function of proteins by reflecting the behavior of amino acids in an aqueous environment. The calculation formula is as follows:
[0022] Step S1.4: Extracting quasi-sequential order (QSO) features.
[0023] QSO captures the global order features in protein sequences, which helps the model identify key sequence patterns related to protein three-dimensional structure and function. The calculation formula is as follows:
[0024] Step S1.5: Extract dipeptide composition (DPC) features.
[0025] DPC calculates the frequency of occurrence of all possible dipeptides (combinations of two consecutive amino acids) in a protein sequence. It reflects the interaction between local amino acid pairs in the sequence, and the calculation formula is as follows:
[0026] Step S2: splicing the five physical and chemical characteristics of the enzyme extracted in step S1 to form an enzyme characteristic sequence.
[0027] Step S3: Construct a model for enzyme sequence feature learning.
[0028] A deep learning network for enzyme sequence feature learning is built based on PyTorch. The network is mainly composed of Transformer and fully connected layers.
[0029] Transformer: The multi-head attention mechanism used in the Transformer model has significant advantages in predicting the solubility of enzymes. The multi-head attention mechanism allows the model to simultaneously focus on features at different positions in the sequence in multiple subspaces, which means that it can capture key information in the sequence at different angles and scales. This mechanism enables the Transformer to simultaneously focus on multiple physical and chemical features and the long-distance dependencies between them when processing complex sequences of enzymes, thereby more comprehensively understanding the important patterns and feature associations in the sequence. This multi-head attention capability enables the model to process information in parallel at different levels, avoiding missing important features, thereby further improving the accuracy and efficiency of solubility prediction. Among them, a head i in the multi-head attention can be described by the following formula:
[0030] Step S4: Input the new enzyme sequence features obtained in step S3 into the fully connected layer to predict the solubility of the enzyme.
[0031] In order to illustrate the superiority of the prediction model constructed by the present invention, RPSP, SKADE, DeepSol, and SWI were selected and used as benchmark models for comparison. The test data set from NESG was selected as the test data set to verify the model. The prediction performance of the model was evaluated using the accuracy (Accuracy, Acc) index and the recall (Recall) index. Among them, the Acc index is mainly used to measure the correct proportion of model predictions, that is, the accuracy reflects the overall correctness of the model in predicting soluble and insoluble enzymes, and the Recall index refers to the proportion of samples that the model correctly identifies as soluble enzymes.
[0032] The results are shown in Table 1. It can be seen from the table that the model is better than other benchmark models in predicting the solubility of enzymes.
[0033] Table 1. Prediction results of NESG dataset Model Accuracy Recall RPSPSKADEDeepSolSWIOur Model 0.4840.4930.5260.5850.596 0.3230.1060.1480.6010.772 Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.
[0034] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for predicting enzyme solubility using a machine learning model. It is characterized in that include: Step S1: constructing a prediction model for predicting enzyme solubility, wherein the prediction model for predicting enzyme solubility comprises: enzyme feature extraction, Transformer, and a fully connected layer; wherein the enzyme feature extraction extracts five physical and chemical characteristics of the enzyme sequence through an algorithm, and these characteristics provide the Transformer model with multi-dimensional and rich sequence information, which is used to improve the accuracy of predicting enzyme solubility; Step S2: using enzyme sequences provided in a public data set as training data to train the prediction model for predicting enzyme solubility; Step S3: Use the trained model to predict the solubility of the enzyme.
2. The method according to claim 1, It is characterized in that The step S2 comprises: Step S2.1: extracting features from enzyme sequences provided in public datasets; Step S2.2: The prediction model for predicting enzyme solubility is trained using the extracted enzyme sequence features, the loss value is calculated during the training process, appropriate indicators are selected for evaluation of the output samples, and the model parameters after the training are saved.
3. The method according to claim 2, It is characterized in that Extract five physicochemical characteristics of enzyme sequences provided in the public dataset, including: Step 1: Design an algorithm to extract the amino acid composition AAC feature, which is used to describe the relative frequency of each amino acid in the protein sequence; Design an algorithm to extract the pseudo amino acid composition PseAAC feature, which describes the amino acid composition information in AAC and combines additional information on the order and physicochemical properties of the amino acids in the sequence; Design an algorithm to extract the amphipathic pseudoamino acids to form APAAC, which describes the hydrophobicity and hydrophilicity of amino acids; Design an algorithm to extract quasi-sequential order QSOs, a feature that emphasizes the relative positions of amino acids and the physicochemical interactions between them; Design an algorithm to extract dipeptide composition DPC, which reflects the interaction of local amino acid pairs in the sequence; Step 2: The five physical and chemical features extracted in step 1 are spliced to form multi-dimensional and rich sequence information; Step 3: Input the sequence features of the enzyme obtained in step 2 into Transformer. Using the attention mechanism of Transformer, the long-distance dependencies between different amino acids can be accurately captured. Step 4: Input the sequence features obtained in step 3 into the fully connected layer to predict the solubility of the enzyme.
4. The method according to claim 3, It is characterized in that Evaluation of model output metrics was used to assess the accuracy of the model in predicting enzyme solubility.
5. Application of the method for predicting enzyme solubility based on a machine learning model as described in any one of claims 1 to 4 in chemicals and pharmaceuticals.