Peptide function prediction method based on artificial intelligence

Abstract

The invention discloses a peptide function prediction method based on artificial intelligence, and belongs to the technical field of biomedical artificial intelligence, and the method comprises the following steps: S1, obtaining a training data set for training an artificial intelligence model; s2, obtaining a prediction model; s3, obtaining a to-be-predicted data set; s4, obtaining a prediction result; and S5, judging the prediction result obtained in the step S4. According to the peptide function prediction method based on artificial intelligence, the peptide chain with the preset function can be efficiently and accurately identified, so that the economic cost of functional peptide screening is remarkably reduced, and the efficiency of the screening process is greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical artificial intelligence technology, and specifically relates to a peptide function prediction method based on artificial intelligence. Background Art

[0002] Peptides are small molecule proteins composed of two or more amino acids connected by peptide bonds. Peptides have many important functions in organisms and are also widely used in many fields such as medicine and industry.

[0003] In the medical field, insulin (extracted from animal pancreas) was first used for medical treatment, changing the treatment of type 1 diabetes. After that, people realized the importance of peptides as key biological mediators, as well as their excellent efficacy, selectivity and low toxicity. At present, the research on new peptide therapies is still progressing steadily.

[0004] In the industrial and food fields, peptides can be used as functional foods. For example, peptides can produce taste or enhance taste. Take umami as an example. Umami, as one of the five basic tastes, was first discovered by Kikunae Ikeda from dried kelp. Umami substances include free amino acids, nucleotides, organic acids, peptides and Maillard reaction products. Among them, umami peptides have attracted much attention due to their good taste and strong processing adaptability. Umami peptides are present in various natural foods, including ham, beef, chicken, mushrooms, peanuts, beans, clams and cod. The most classic umami peptide is beef peptide extracted from beef broth treated with papain. The discovery and identification of new umami peptides from food-derived proteins is gaining more and more interest.

[0005] As the number of peptide sequences continues to grow in the post-genomic era, experimentally identifying the function of peptides remains time-consuming and costly. Therefore, being able to quickly and accurately identify the function of peptides is critical for developing pharmaceuticals and advancing nutritional research. Summary of the invention

[0006] The purpose of the present invention is to provide a peptide function prediction method based on artificial intelligence, which can efficiently and accurately identify peptide chains with predetermined functions, thereby significantly reducing the economic cost of functional peptide screening and greatly improving the efficiency of the screening process.

[0007] To achieve the above object, the present invention provides a peptide function prediction method based on artificial intelligence, comprising the following steps:

[0008] S1. Obtain a training data set for training an artificial intelligence model;

[0009] S2. Inputting the training data set obtained in S1 into a pre-trained artificial intelligence model or an unpre-trained artificial intelligence model for training, and evaluating the trained artificial intelligence model to obtain a prediction model focusing on a specific function;

[0010] S3, obtaining a data set to be predicted;

[0011] S4, inputting the data set to be predicted obtained in S3 into at least two prediction models in a serial / parallel manner for prediction to obtain a prediction result, wherein the prediction result is used to reflect whether the data set to be predicted has the expected function;

[0012] S5. Determine the prediction results obtained in S4. When all the prediction results are higher than a preset threshold, determine that the peptide chain has the expected function.

[0013] Preferably, in S1, the artificial intelligence model includes a deep learning model and a machine learning model, the deep learning model includes a transformer model, a BERT-based model, a deep neural network, and a convolutional neural network, and the machine learning model includes logistic regression, support vector machine, random forest, gradient boosting tree, K-nearest neighbor algorithm, linear discriminant analysis, decision tree, naive Bayes or perceptron.

[0014] Preferably, the deep learning model and the machine learning model both include classification algorithms and regression algorithms.

[0015] Preferably, in S1, existing literature and public databases of positive peptides and negative peptides are systematically consulted to collect functional positive peptide and negative peptide data information, and data feature expansion is performed on the collected data information to generate a training data set for training an artificial intelligence model.

[0016] Preferably, in S3, an initial peptide chain is obtained, and biological activity prediction, toxicity prediction and peptide property prediction are performed on the initial peptide chain. The initial peptide chain that passes the biological activity prediction, toxicity prediction and peptide property prediction is identified as the peptide chain to be predicted, and data feature expansion is performed on the peptide chain to be predicted to generate a data set to be predicted.

[0017] Preferably, data feature expansion refers to feature mining, i.e. mining amino acid composition, dipeptide composition, normalized Moreau-Broto autocorrelation, composition-transition-distribution, quasi-sequence coupling number, pseudo-amino acid composition, weighted pseudo-amino acid composition and basic descriptors.

[0018] Preferably, the method for obtaining the initial peptide chain comprises the following:

[0019] The method based on material research: search a species in the protein database and obtain all its protein structures, then hydrolyze the obtained protein structures with enzymes, and obtain the initial peptide chain after the hydrolysis is completed;

[0020] Method based on the importance of peptides: several important amino acids in the peptide chain to be predicted by the prediction model in S2 are obtained through the Shap method, the characteristic importance of the model or the weight distribution of the model, and the amino acids with higher frequency at the C-terminus or N-terminus are obtained through the C-terminus and N-terminus frequency analysis. The important QSAR features in the prediction model in S2 are obtained through the Shap method, the characteristic importance of the model or the weight distribution of the model. The prediction peptides are designed in a targeted manner through the above indicators to obtain the initial peptide chain;

[0021] The method of random amino acid arrangement: the initial peptide chain is obtained by randomly arranging amino acids.

[0022] Preferably, in S4, the series / parallel connection includes a series model structure, a parallel model structure and a series-parallel mixed model structure.

[0023] Preferably, S5 also includes experimental verification of the peptide chain having the expected function.

[0024] Preferably, in S5, the intended functions include alpha-amylase inhibitor, anti-amnesia, ACE inhibitor, celiac toxicity, antibacterial, antithrombotic, immunomodulatory, opioid, irritant, immunostimulatory, neuropeptide, regulatory, anticancer, antioxidant, umami, bitter, sweet or sour.

[0025] Therefore, the present invention adopts the above-mentioned peptide function prediction method based on artificial intelligence. Compared with the prior art, the present invention has the following significant beneficial effects:

[0026] (1) This invention innovatively proposes a deep learning model based on sequence and a large number of QSAR (Quantitative Structure-Activity Relationship) features;

[0027] (2) The present invention uses computational simulation technology to replace a large number of laboratory wet experiments to obtain quantitative indicators of functional peptides. This not only significantly reduces experimental costs, but also greatly reduces the consumption of human resources, thereby improving the efficiency of research work;

[0028] (3) This invention proposes a new model that can not only effectively determine the functionality of peptides by predicting the binding energy of molecular docking, but also quantify the strength of peptide functions. This innovative method provides more detailed and quantitative data support for subsequent research work;

[0029] (4) The model proposed in the present invention is not limited to the prediction of a specific type of functional peptides, but is applicable to all scenarios involving ligand-receptor binding effects, and has wide applicability and flexibility;

[0030] (5) The present invention can achieve extremely high computing speed while maintaining high precision, greatly improving processing efficiency.

[0031] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flowchart of an embodiment of a peptide function prediction method based on artificial intelligence of the present invention;

[0033] Figure 2 A data feature expansion flow chart of a peptide function prediction method based on artificial intelligence of the present invention;

[0034] Figure 3 A flow chart of prediction model selection for a peptide function prediction method based on artificial intelligence of the present invention;

[0035] Figure 4 A flowchart of peptide chain prediction processing of a peptide function prediction method based on artificial intelligence of the present invention;

[0036] Figure 5 A prediction flow chart of a peptide function prediction method based on artificial intelligence of the present invention;

[0037] Figure 6 This is a graph of ACE inhibition rate in Example 2 of a peptide function prediction method based on artificial intelligence of the present invention;

[0038] Figure 7 This is a radar chart of umami peptides in Example 3 of a peptide function prediction method based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0041] Embodiment 1

[0042] like Figure 1-Figure 5 As shown, a peptide function prediction method based on artificial intelligence of the present invention comprises the following steps:

[0043] S1. Obtain training data set for training artificial intelligence model:

[0044] Systematically review existing literature and public databases of positive peptides and negative peptides, collect or calculate positive peptides and negative peptides with corresponding functions and their functional evaluation indicators (e.g., inhibition rate, MIC value, IC50 value, and minimum binding energy, etc.). Perform data feature expansion on the collected peptide chains to generate a training data set for training artificial intelligence models.

[0045] S2. Obtain prediction model:

[0046] The training data set obtained in step S1 is input into a pre-trained artificial intelligence model or an artificial intelligence model that has not been pre-trained for further training. A pre-trained artificial intelligence model means that the model has gone through an initial training phase before it is officially used. For example, before training a model for umami peptides, the model has been trained using a training data set consisting of other flavor peptides such as bitter peptides and sour peptides. The purpose of pre-training is to give the model a better initial weight so that in the subsequent retraining process, the model can reach a fitting state faster. An artificial intelligence model that has not been pre-trained means that the model has not undergone any initial training before starting to use the model.

[0047] Both pre-trained models and non-pre-trained models are artificial intelligence models. The artificial intelligence models used include deep learning models and machine learning models. Deep learning models include transformer models, BERT-based models, deep neural networks, and convolutional neural networks. Machine learning models include logistic regression, support vector machines, random forests, gradient boosting trees, K-nearest neighbor algorithms, linear discriminant analysis, decision trees, naive Bayes, or perceptrons.

[0048] Both deep learning models and machine learning models include classification algorithms and regression algorithms. Choose the appropriate algorithm based on the prediction target. If the task is to determine whether a peptide has a certain effect, regardless of the degree of the effect, a classification algorithm should be used, for example, to determine whether a peptide has antibacterial activity. If the task requires evaluating the degree of efficacy of the peptide, that is, not only knowing whether the peptide is effective, but also knowing how strong its effect is, a regression algorithm should be used, for example, to predict the strength of a peptide's antibacterial activity. The choice of algorithm depends on the desired prediction value (Y). Specifically as follows:

[0049] Classification algorithm: The peptides (data) provided for training are negative peptides (data) or positive peptides (data), that is, they have the function (1) or do not have the function (0). When the model trained by this algorithm is used to predict the expected peptide, the predicted value will be between 0 and 1. Peptides with predicted values ​​higher than a certain threshold are considered to have the function (1), and peptides with predicted values ​​lower than a certain threshold are considered not to have the function (0).

[0050] Regression algorithm based on molecular docking binding energy: the peptides (data) provided for training are negative peptides (data) or positive peptides (data), and then these negative peptides or positive peptides are molecularly docked and their binding energy is recorded, and the binding energy is used as a prediction indicator.

[0051] Regression algorithm based on functional evaluation index: The peptides (data) provided for training are negative peptides (data) or positive peptides (data), but the positive data has a functional range. For example, 0-1, the inhibition rate is 0 as a negative peptide (completely without the function), and the inhibition rate is 1 as the best peptide; between 0 and 1, it is considered to be effective. Taking antimicrobial peptides as an example, 0 represents no antibacterial effect at all, and any value greater than 0 and less than or equal to 1 represents antibacterial effect. At this time, the inhibition rate or IC50 value can be used as training data or prediction data. When the model trained by this algorithm is used to predict the expected peptide, the prediction value can be between 0 and 1, and the prediction value is considered to be gradually weakened from high to low. Similarly, for some peptides with adverse effects, it is assumed that the functional range is between -1 and 1, and a value of -1 is considered to have complete side effects, a value between -1 and 0 is considered to have side effects, a value of 0 is considered to have no effect, a value between 0 and 1 is considered to have a positive effect, and a value of 1 is considered to have a complete positive effect.

[0052] Evaluate the trained model to obtain a prediction model that focuses on a specific function. During the training process, many models will be involved in the training, and finally the best one or more models will be selected through model evaluation for the subsequent prediction process. The optional specific model evaluation indicators are as follows:

[0053] Accuracy is a basic indicator for evaluating the performance of a classification model. It measures the model's ability to correctly predict samples. The calculation formula for accuracy is as follows:

[0054] ACC=(TP+TN) / (TP+TN+FP+FN)

[0055] Among them, TP represents the number of correctly predicted positive samples, TN represents the number of correctly predicted negative samples, FP represents the number of negative samples incorrectly predicted as positive, also known as false positives, and FN represents the number of positive samples incorrectly predicted as negative, also known as false negatives.

[0056] The F1 score is an important evaluation indicator in classification problems. It is the harmonic mean of precision and recall. The F1 score is used to measure the accuracy and robustness of the model. The definitions of precision and recall are as follows:

[0057] Precision: The proportion of samples that are actually positive among all samples predicted by the model to be positive. The formula is:

[0058]

[0059] Among them, TP represents the number of correctly predicted positive samples, and FP represents the number of negative samples that are incorrectly predicted as positive.

[0060] Recall rate: The proportion of samples that are actually positive that the model correctly predicts as positive, also expressed as sensitivity or True Positive Rate (TPR). The formula is:

[0061]

[0062] Among them, TP represents the number of correctly predicted positive samples, and FN represents the number of positive samples that are incorrectly predicted as negative samples.

[0063] The calculation formula of F1 score is:

[0064]

[0065] Specificity: The proportion of samples that are actually negative among all samples predicted by the model to be negative. The formula is as follows:

[0066]

[0067] The Receiver Operating Characteristic (ROC) curve is a graph widely used to evaluate the performance of classification models. It is used to show the relationship between the sensitivity and specificity of a classifier at different thresholds.

[0068] AUC (Area Under the ROC Curve) is the area under the ROC curve, which is used to quantify the model's ability to distinguish between positive and negative classes.

[0069] Matthews Correlation Coefficient (MCC) is an indicator used to evaluate the prediction quality of a binary classification model, which takes into account all four possible classification results: TP, FP, FN, and TN. The calculation formula of MCC is as follows:

[0070]

[0071] Mean Squared Error (MSE) Mean Squared Error (MSE) is the average of the squares of the differences between the predicted values ​​and the true values, and the formula is as follows:

[0072]

[0073] Among them, y i is the ith true value, is the ith predicted value and n is the number of data points.

[0074] Root Mean Squared Error (RMSE) is a derivative of the mean square error (MSE). It provides an error measure in the same unit as the original data by taking the square root of the MSE. The formula is as follows:

[0075]

[0076] Mean Absolute Error (MAE) is the average of the absolute values ​​of the differences between the predicted values ​​and the true values, and the formula is as follows:

[0077]

[0078] Coefficient of Determination (R 2 ) is a statistic that measures the goodness of fit of a regression model. It indicates the proportion of the total variability of the response variable explained by the model. The formula is as follows:

[0079]

[0080] in, is the average of all true values, is the sum of squares for residuals, that is, the sum of squares of the differences between the model predictions and the true values, is the Total Sum of Squares, which is the sum of squares of the differences between the true value and the mean of the response variable.

[0081] S3. Obtain the data set to be predicted:

[0082] The methods for obtaining the initial peptide chain by simulating enzymatic hydrolysis and other methods include:

[0083] The method based on material research uses UniProtKB or other protein databases to retrieve protein sequence information of a specific species and obtain all its protein structures. The appropriate protease is selected for hydrolysis, and the initial peptide chain can be obtained after the hydrolysis is completed.

[0084] The method based on the importance of peptides, after model training, selects single amino acids or multiple amino acids with higher prediction weights for the final model and randomly arranges them to obtain the initial peptide chain; on this basis, the amino acid sequence regularity of positive peptides can be explored, such as the frequency of amino acids appearing at the N-terminus or C-terminus, and possible initial peptide chains can be designed in a targeted manner.

[0085] The method of randomly arranging amino acids to explore all possible peptide chain combinations. For example, peptide chains of different lengths such as dipeptides, tripeptides, and tetrapeptides are generated for subsequent research.

[0086] After obtaining the initial peptide chain in the above manner, the initial peptide chain is subjected to biological activity prediction, toxicity prediction and peptide property prediction, and peptides with specific properties and potential application value are screened out, such as peptides with biological activity above a certain threshold, non-toxicity and good water solubility. After the initial peptide chain is screened as above, the peptide chain to be predicted is obtained, and the data feature of the peptide chain to be predicted is expanded to generate a data set to be predicted.

[0087] The data feature expansion in S1 and S3 includes the following steps:

[0088] 1. For the training data set: positive peptides and negative peptides of corresponding functions are obtained from various databases or literatures. At this time, the data are all peptide chain sequences (composed of 20 amino acids). For the data set to be predicted: the initial peptide chain is obtained by various methods such as simulated enzymatic hydrolysis, and the initial peptide chain is predicted through biological activity prediction, toxicity prediction, and peptide property prediction. The initial peptide chain sequence that passes all predictions is identified as the peptide chain sequence to be predicted.

[0089] 2. Expand the data features of the obtained peptide chain sequence, and the expanded content includes:

[0090] i. The proportion of amino acids in the peptide chain. Specifically, the proportion of the twenty amino acids in the peptide chain. Taking AH as an example, amino acid A and amino acid H each account for 50%, and the remaining 18 amino acids account for 0% in the peptide chain;

[0091] ii. The proportion of dipeptides in the peptide chain. Specifically, among the 400 dipeptides, taking ACH as an example, dipeptide AC and dipeptide CH each account for 50%, while the remaining 398 dipeptides account for 0% in the peptide chain.

[0092] iii. Characteristics such as hydrophobicity, flexibility, polarizability, free energy, self-fusion area, volume and steric hindrance.

[0093] In order to comprehensively describe the characteristics of peptide chains, a variety of calculation methods were used in the process of data feature expansion. These calculation methods include amino acid composition (AAC), dipeptide composition (DIP), normalized Moreau-Broto auto-correlation (MBA), composition-transition-distribution (CTD), quasi-sequence order coupling number (QSO), pseudo amino acid composition (PAAC), weighted pseudo amino acid composition (APAAC) and basic descriptors.

[0094] 3. After data feature expansion, all acquired features are retained. The data set originally containing only peptide sequence data is transformed into a data set containing both peptide sequence data and QSAR feature data. The peptide sequence and QSAR features together constitute the input parameter (X).

[0095] S4. Obtain prediction results:

[0096] The artificial intelligence models are combined to obtain a series model structure, a parallel model structure and a series-parallel hybrid model structure.

[0097] The tandem model structure means that the data set to be predicted first passes through one model and then passes through one or more models to obtain the result, and the models are in tandem. Generally speaking, the first model is a classification model, which can quickly distinguish the effectiveness of the predicted peptide, and the second and subsequent models are regression models, which can accurately distinguish the efficiency of the predicted peptide.

[0098] The parallel model structure means that the data set to be predicted passes through one model while passing through another one or more models, and finally the prediction results of the two or more models are compared and merged to generate a prediction result. The two or more models can be classification models and regression models at the same time, or part of them can be classification models and part of them can be regression models.

[0099] The series-parallel hybrid model structure refers to combining the above two methods. The data set to be predicted can pass through one series path while passing through another series path. Finally, the prediction results are combined according to certain weights or in a certain way to generate the final prediction results.

[0100] After the data set to be predicted is predicted in the above manner, a prediction result is obtained, and the prediction result is used to determine whether the peptide chain has the expected function.

[0101] S5. Determine the prediction result obtained in S4:

[0102] After a set of predicted data passes through a model, a set of predicted values ​​will be generated. The predicted value represents the confidence level. For different tasks, the higher or lower the predicted value represents different meanings. Taking the 01 binary classification task as an example, the closer the predicted value is to 1, the more likely the predicted peptide is to be positive. After passing through multiple models, multiple sets of predicted values ​​will be generated; the multiple sets of predicted values ​​generated can be used to obtain the final predicted value through the weighted summation method, or they can be determined by the fact that each single set of predicted values ​​is higher than different or the same values.

[0103] A threshold is set, and when all predicted values ​​are higher than the threshold, the peptide chain is considered to have the expected function. After the above screening, peptide chains with the expected function are obtained, and these peptide chains are verified by wet experiments to screen out functional peptides that meet the conditions.

[0104] Embodiment 2

[0105] like Figure 6 As shown, in this embodiment, a functional peptide of an ACE inhibitory peptide is obtained. First, peptides with ACE inhibitory activity are screened from the BioPep database to ensure that they contain only 20 basic amino acids and exclude irregular peptides, and finally 1089 positive peptides are obtained. Peptides with a length of no more than 30 amino acids, which have been reviewed and whose functions do not include lowering blood pressure or vasodilation, are selected from the UniProtKB database, and 1089 are randomly sampled as negative peptides. The collected 1089 positive peptides and 1089 negative peptides are combined and data features are expanded, and the steps include:

[0106] 1. The proportion of amino acids in the peptide chain. Specifically, the proportion of the twenty amino acids in the peptide chain. Taking AH as an example, amino acid A and amino acid H each account for 50%, and the remaining 18 amino acids account for 0% in the peptide chain;

[0107] 2. The proportion of dipeptides in the peptide chain. Specifically, the proportion of the 400 dipeptides in the peptide chain, taking ACH as an example, dipeptide AC and dipeptide CH each account for 50%, while the remaining 398 dipeptides account for 0% in the peptide chain;

[0108] 3. The calculation methods used in the process of data feature expansion include normalized Moreau-Broto Auto-Correlation (MBA), composition-transition-distribution (CTD), quasi-sequence order coupling number (QSO), pseudo amino acid composition (PAAC), weighted pseudo amino acid composition (APAAC) and basic descriptor features.

[0109] The resulting training dataset contains both peptide sequences and these expanded physical and chemical indicators.

[0110] A pre-trained deep learning BERT model and an unpre-trained machine learning random forest model were used. The deep learning BERT model was pre-trained with functional peptides such as anti-cancer peptides, neuropeptides, and anti-tuberculosis peptides, and the model was improved so that it can consider both peptide sequences and QSAR features. In fact, the above two models were screened out from 9 models. Models that can only consider the QSAR features (table) of peptides include logistic regression models, random forests, XGBoost, SVC, DNN, CNN (the ratio of training set to validation set is 8:2); models that can simultaneously consider the sequence (seq) and QSAR features (table) of peptides include BERT-seq, BERT-seq+CNN-table, BERT-seq+DNN-table (using five-fold validation). The evaluation scores of the models are as follows:

[0111] Table 1 Model evaluation score table

[0112]

[0113] According to the evaluation scores of the models, bert-seq+dnn-table achieved the best accuracy, precision, and sensitivity when considering both peptide sequence and QSAR features, so the bert-seq+dnn-table model was finally selected as the final prediction model. Random forest-table achieved the best accuracy, precision, and specificity when only considering QSAR features, so the random forest-table model was selected as the verification model.

[0114] All reviewed proteins of Pleurotus ostreatus were selected from the UniprotKB database, and a total of 13 proteins were obtained. PeptideCutter was used to simulate enzymatic hydrolysis of the 13 proteins. Chymotrypsin, pepsin and trypsin were used in the simulated enzymatic hydrolysis to remove peptides after single amino acids, and finally 420 peptides were obtained.

[0115] After that, PeptideRanker was used to predict the biological activity of 420 peptides, and 52 peptides with predicted biological activity values ​​higher than 0.5 were selected. ToxinPred was then used to predict the toxicity of these 52 peptides, and the prediction results showed that all 52 peptides were non-toxic. The peptide properties of these 52 peptides were predicted again, and the results showed that all 52 peptides met the preset requirements.

[0116] The data to be predicted are input into the model in parallel. The bert-seq+dnn-table model is selected as the prediction model, and the random forest-table model is selected as the verification model. The peptide chains and QSAR features of 52 peptides are put into the two models at the same time to obtain the prediction values ​​of the two models.

[0117] The peptide chains with predicted values ​​greater than 0.5 in both models were selected, and the top five peptides were selected as ACE inhibitory peptides according to the predicted values. The five ACE inhibitory peptides selected were verified by wet experiments to prove their effectiveness.

[0118] Embodiment 3

[0119] like Figure 7As shown, in this embodiment, the functional peptide to be obtained is an umami peptide. The collection of positive peptides was obtained through the database BioPep and published literature. The screening range in BioPep was peptides containing umami or umami enhancing in activity (Activity), and finally 337 positive peptides were obtained. The acquisition of negative peptides was obtained through BioPep and published literature. The screening range in BioPep was non-umami peptides (i.e., bitter, sour, and astringent peptides), and 337 data were randomly sampled as negative peptides. Afterwards, by combining these 337 positive peptides and 337 negative peptides and expanding the data features, the training data set finally formed not only contains the peptide sequence, but also includes these expanded physical and chemical indicators.

[0120] Molecular docking was performed on the above 337 positive peptides. Receptor processing: Since the crystal structures of human type 1 taste receptor member 1 (T1R1) and type 1 taste receptor member 3 (T1R3) are still unclear, the three-dimensional (3D) structure of the umami receptor subunit T1R1 / T1R3 was constructed using AlphaFold3. The amino acid sequences of T1R1 and T1R3 are from the UniProt database (T1R1: Q7RTX1; T1R3: Q7RTX0). Ligand processing: The three-dimensional structure of the ligand was generated using pymol software or the rdkit package in python or using Chem3D, and hydrogenation and removal of small molecule ligands were performed.

[0121] Molecular docking: The processed ligands were docked with T1R1 / T1R3 by AutoDock Vina and the minimum binding energy was taken as the prediction value of the training set. 337 negative peptides were considered impossible to bind to T1R1 / T1R3, and their molecular docking binding energy was set to 1024 (generally, the binding energy of molecules that can be docked is negative, and 1024 means that they cannot be docked).

[0122] It should be noted that whether it is an umami peptide is still marked, umami peptide is 1, and non-umami peptide is 0. The purpose is to quickly and accurately retain umami peptides during preliminary screening to improve efficiency.

[0123] A pre-trained deep learning BERT model and an untrained XGBoost model were used. In fact, the above two models were selected from 9 models. These two models were tested by R 2 , MAE, MSE and RMSE evaluations obtained good results.

[0124] Through random arrangement of amino acids, 400 dipeptides and 8,000 tripeptides were initially constructed; then, by analyzing the frequency ratio of N-terminal and C-terminal amino acids and the importance of shap, 1,000 peptides were selected for molecular docking binding energy prediction, which can reflect whether they have umami taste.

[0125] Then, the 1000 peptides were used for biological activity prediction by PeptideRanker, and 496 peptides with predicted biological activity values ​​higher than 0.5 were selected for subsequent steps. Then, the toxicity of these 496 peptides was predicted by ToxinPred, and the prediction results showed that all 496 peptides were non-toxic. Then, the 496 peptides were predicted by peptide properties, showing that all 496 peptides met the requirements.

[0126] The dataset to be predicted was input into the model in series. The XGBoost model was selected as the preliminary screening model, and the bert-seq+dnn-table model was used as the final prediction model. The QSAR features of the 496 peptides were first passed through the XGBoost model to preliminarily determine whether they have umami taste, and then the bert-seq+dnn-table model was used to quantitatively determine the concentration of umami taste.

[0127] The molecular docking binding energies predicted by the BERT-Seq+DNN-TABLE model were arranged in reverse order, with smaller values ​​representing tighter binding, and the top five peptides were selected for experimental verification.

[0128] It is worth noting that the contents not elaborated in detail in the present invention are all prior art and are well known to those skilled in the art.

[0129] Therefore, the present invention adopts the above-mentioned artificial intelligence-based peptide function prediction method, which can efficiently and accurately identify peptide chains with predetermined functions, thereby significantly reducing the economic cost of functional peptide screening and greatly improving the efficiency of the screening process.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A peptide function prediction method based on artificial intelligence, characterized in that: The following steps are involved: S1. Obtain a training data set for training an artificial intelligence model; S2. Inputting the training data set obtained in S1 into a pre-trained artificial intelligence model or an unpre-trained artificial intelligence model for training, and evaluating the trained artificial intelligence model to obtain a prediction model focusing on a specific function; S3, obtaining a data set to be predicted; S4, inputting the data set to be predicted obtained in S3 into at least two prediction models in a serial / parallel manner for prediction to obtain a prediction result, wherein the prediction result is used to reflect whether the data set to be predicted has the expected function; S5. Determine the prediction results obtained in S4. When all the prediction results are higher than a preset threshold, determine that the peptide chain has the expected function.

2. The method for predicting peptide function based on artificial intelligence according to claim 1, characterized in that: In S1, the artificial intelligence model includes a deep learning model and a machine learning model. The deep learning model includes a transformer model, a BERT-based model, a deep neural network, and a convolutional neural network. The machine learning model includes logistic regression, support vector machine, random forest, gradient boosting tree, K-nearest neighbor algorithm, linear discriminant analysis, decision tree, naive Bayes or perceptron.

3. The method for predicting peptide function based on artificial intelligence according to claim 2, characterized in that: Both the deep learning model and the machine learning model include classification algorithms and regression algorithms.

4. The method for predicting peptide function based on artificial intelligence according to claim 1, characterized in that: In S1, the existing literature and public databases of positive peptides and negative peptides are systematically consulted to collect functional positive peptide and negative peptide data information, and the data features of the collected data information are expanded to generate a training data set for training artificial intelligence models.

5. The method for predicting peptide function based on artificial intelligence according to claim 4, characterized in that: In S3, an initial peptide chain is obtained, and biological activity prediction, toxicity prediction and peptide property prediction are performed on the initial peptide chain. The initial peptide chain that passes the biological activity prediction, toxicity prediction and peptide property prediction is identified as the peptide chain to be predicted, and data feature expansion is performed on the peptide chain to be predicted to generate a data set to be predicted.

6. The method for predicting peptide function based on artificial intelligence according to claim 5, characterized in that: Data feature expansion refers to feature mining, namely mining amino acid composition, dipeptide composition, normalized Moreau-Broto autocorrelation, composition-transition-distribution, quasi-sequence coupling number, pseudo-amino acid composition, weighted pseudo-amino acid composition and basic descriptors.

7. The method for predicting peptide function based on artificial intelligence according to claim 5, characterized in that: Methods for obtaining the initial peptide chain include the following: The method based on material research: search a species in the protein database and obtain all its protein structures, then hydrolyze the obtained protein structures with enzymes, and obtain the initial peptide chain after the hydrolysis is completed; Method based on the importance of peptides: several important amino acids in the peptide chain to be predicted by the prediction model in S2 are obtained through the Shap method, the characteristic importance of the model or the weight distribution of the model, and the amino acids with higher frequency at the C-terminus or N-terminus are obtained through the C-terminus and N-terminus frequency analysis. The important QSAR features in the prediction model in S2 are obtained through the Shap method, the characteristic importance of the model or the weight distribution of the model. The prediction peptides are designed in a targeted manner through the above indicators to obtain the initial peptide chain; Method of random amino acid arrangement: the initial peptide chain is obtained by random arrangement of amino acids.

8. The method for predicting peptide function based on artificial intelligence according to claim 1, characterized in that: In S4, series / parallel connection includes a series model structure, a parallel model structure, and a series-parallel mixed model structure.

9. The method for predicting peptide function based on artificial intelligence according to claim 1, characterized in that: S5 also includes experimental verification of the peptide chain having the expected function.

10. The method for predicting peptide function based on artificial intelligence according to claim 9, characterized in that: In S5, the intended functions include alpha-amylase inhibitor, anti-amnestic, ACE inhibitor, celiac toxicity, antimicrobial, antithrombotic, immunomodulatory, opioid, irritant, immunostimulatory, neuropeptide, regulatory, anticancer, antioxidant, umami, bitter, sweet, or sour.

Images