A deep learning algorithm to predict stability of polypeptides in different blood environments

Abstract

The application discloses a kind of deep learning algorithm for predicting polypeptide in different blood stabilities, specifically includes: from public database systematically collects and organizes about polypeptide in human body, mouse body and in-vivo and in-vitro blood half-life experimental data and after strict data cleaning and verification process, construct high-quality data set as the basis of model training.Secondly create a multi-modal prediction model PepMSND, the model considers four key dimensional features of polypeptide, namely zero-dimensional polypeptide physicochemical property, one-dimensional polypeptide SMIELS code, two-dimensional polypeptide molecular graph, three-dimensional polypeptide 3D structure.In order to effectively process these different dimensional features, PepMSND integrates KAN, Transformer, GAT and SE (3) -Transformer and other advanced neural network components.The algorithm of the application provides an effective technical solution for the high cost and high time-consuming problem of plasma stability test in polypeptide drug research and development, thereby accelerating the early development process of polypeptide drug.

Description

Technical Field

[0001] The present invention belongs to the technical field of polypeptide drug research and development, and specifically relates to a deep learning algorithm for predicting the stability of polypeptides in different blood environments. Background Art

[0002] Peptide and protein drugs, due to their unique advantages, occupy a vital position in the development of the biopharmaceutical field. To date, nearly 80 peptide drugs have been approved for clinical use in the treatment of various diseases. However, despite the remarkable clinical success of peptide drugs, the number of new peptide drugs entering clinical trials has not shown the same growth trend. Peptides are highly susceptible to hydrolysis by proteases in the body, especially in plasma. This instability severely hinders the conversion of bioactive peptides into drug leads. Therefore, the stability of peptides in plasma has become an important consideration in clinical applications. Accurately predicting their stability in plasma is crucial for the development of peptide drugs.

[0003] Traditionally, the prediction of peptide stability has relied primarily on experimental methods, such as plasma stability tests and enzymatic hydrolysis assays. While these methods can accurately assess peptide stability, they are time-consuming and costly, making them unsuitable for high-throughput screening and large-scale studies. Currently, machine learning algorithms exist for predicting peptide stability in plasma. However, these models often overlook the significant differences in test species and environments when constructing training data. This significant variability can lead trained models to incorrectly select inappropriate lead compounds, ultimately leading to failure in subsequent peptide drug development. Furthermore, previous models for predicting peptide stability or other properties have generally tended to characterize peptides using relatively simple, low-dimensional features as input for model training. However, a significant limitation of these methods is that they often overlook the unique 3D structural information of peptides. Given the current challenges in predicting peptide stability in blood, a multimodal computational approach that comprehensively considers different blood species and experimental conditions is urgently needed. Summary of the Invention

[0004] In response to the above problems, the purpose of the present invention is to provide a deep learning algorithm for predicting the stability of polypeptides in different blood environments.

[0005] The specific technical solutions are as follows:

[0006] A deep learning algorithm for predicting peptide stability in different blood environments includes the following steps:

[0007] Step 1: Obtain a dataset of peptide stability in plasma;

[0008] Step 2: Clean and organize the data collected in step 1. Data that do not meet the requirements include: 1. Excluding peptides with missing half-life data or uncertain half-life data; 2. Excluding peptides for which the corresponding sequence information cannot be determined; 3. Excluding peptides with complex modifications that cannot be effectively represented by SMILES codes, such as peptides modified with polyethylene glycol or bound to macromolecular carriers such as proteins and lipids.

[0009] Step 3: After completing data cleaning and organization, SMILES encoding representation was added to all peptides using ChemDraw and RDKit.

[0010] Step 4: After completing step 3, use tools such as AlphaFold2, HighFold, and RDKit to build a three-dimensional structural model for each peptide.

[0011] Step 5: Use RDKit to calculate the physicochemical properties of the peptide, and use the random forest algorithm to screen out the top 140 features that have a significant impact on the blood stability of the peptide and retain them.

[0012] Step 6: After completing all the previous processing steps, the data is classified into stable and unstable categories based on whether the half-life of the peptide in the blood is greater than 1 hour. The data is also divided into training and test sets for model training and testing.

[0013] Step 7: For the preprocessed data, a multimodal model PepMSND is constructed for training. The model includes: a KAN module for receiving zero-dimensional features, a Transformer module for processing one-dimensional features, a GAT module for analyzing two-dimensional features, and a SE(3)-Transformer module for parsing three-dimensional features.

[0014] Step 8: For the outputs of the four modules (KAN, Transformer, GAT, SE(3)-Transformer) of the multimodal model PepMSND, first multiply each by a learnable weight W, and then concatenate the four weighted outputs.

[0015] Step 9: The spliced ​​output is used as the input of the KAN model, and the characteristic information of the species and experimental environment is added. Finally, a one-dimensional output is generated through dimensionality reduction processing, which represents the confidence level of the peptide stability in the blood.

[0016] Step 10: After model training, the performance of the multimodal model PepMSND is evaluated using the test set. Evaluation metrics include ACC (Accuracy), Precision, Recall, F1-score, AUC (Area Under the Curve), and MCC (Matthews Correlation Coefficient).

[0017] Furthermore, the peptide stability dataset in plasma in step 1 was collected and searched from public databases such as PEPlife, THPdb, DrugBank, and PubMed using “peptide”, “half-life”, and “blood” as keywords;

[0018] Furthermore, step four specifically includes: for linear peptides containing only natural amino acids and no modifications, AlphaFold2 is used to generate the 3D structure; and for cyclic peptides containing only natural amino acids and no modifications, HighFold is used for generation. Due to the limitations of current protein prediction tools for peptides containing complex modifications, RDKit (version 2023.3.2) was used to generate the 3D structure of each cyclic peptide. First, 5000 conformations were generated for each peptide segment, and then the UFF force field was used to optimize the structure of the 5000 conformations. Finally, the structure with the lowest potential energy was selected as the final conformation of the peptide.

[0019] Furthermore, step five specifically involves first removing descriptors that maintain a constant value across peptide molecules to reduce redundant information. Subsequently, a random forest model is constructed based on the remaining descriptors. By evaluating the feature importance of the random forest model, the top 140 most influential molecular descriptors are selected as input features. Furthermore, species and experimental environment information are introduced as additional input features via one-hot encoding.

[0020] Furthermore, step seven specifically includes: 1. Using the 140 molecular descriptors obtained in step five and the introduced species and experimental environment information, a total of 142 features, are input into the KAN model with a size of N×142. 2. The SMILES codes of all peptides are expanded to a length of 1024 and input into the Transformer model with a size of N×1024. 3. For the input of GAT, the DGL python package is used to generate a molecular graph and input it into GAT. 4. The coordinates are extracted from the PDB file of each peptide as the input of SE(3)-transformer.

[0021] Furthermore, step eight specifically includes: adjusting the output size of the four modules, KAN, Transformer, GAT, and SE(3)-Transformer, to N×128. Then, the outputs of these four modules are multiplied by the corresponding learnable weights W1, W2, W3, and W4 respectively. Then, the weighted outputs are concatenated to form a matrix of size N×512, and finally, this matrix is ​​input into the subsequent KAN model.

[0022] Furthermore, step nine specifically involves passing the N×512 matrix obtained in step eight through a KAN linear layer to reduce its dimensionality to an N×8 matrix. Species and experimental environment information is then incorporated using one-hot encoding and concatenated onto the N×8 matrix, resulting in an N×10 matrix. Finally, this dimensionality reduction process generates a one-dimensional output representing the confidence level of the peptide's stability in blood. The confidence threshold is set to 0.5, with a confidence level greater than 0.5 considered stable and a confidence level less than 0.5 considered unstable.

[0023] Furthermore, in step 10, ACC, Precision, Recall, F1-score, AUC, and MCC are used to evaluate the performance of the model. The specific calculation formula is as follows:

[0024]

[0025]

[0026] The true positives TP, true negatives TN, false positives FP, and false negatives FN are used to represent the counts of each category respectively.

[0027] The beneficial effects of the present invention are: on the one hand, when constructing the data set, the polypeptide stability data are marked and classified in detail according to the species and experimental environment, which solves the problem of ignoring the differences in test species and test environment in the original method. On the other hand, the present invention adopts the multimodal model PepMSND, which integrates the four-dimensional information of the polypeptide (zero-dimensional features, one-dimensional features, two-dimensional features and three-dimensional features). In this way, the factors affecting the stability of polypeptides in the blood can be understood and evaluated from an all-round and multi-dimensional perspective, thereby improving the accuracy and generalization ability of the prediction. The algorithm designed by the present invention to predict the stability of polypeptides in different blood environments will greatly promote the research speed of polypeptide drugs and reduce the research and development costs of such drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The flowchart of the present invention is a specific implementation of the prediction of polypeptide stability in different blood conditions. DETAILED DESCRIPTION

[0029] The present invention will be described in detail below with reference to the accompanying drawings, providing a clear and complete description of the technical solutions in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0030] like Figure 1 As shown, a deep learning algorithm for predicting the stability of peptides in different blood environments includes the following steps:

[0031] Step 1: Collect peptide stability datasets in plasma from public databases such as PEPlife, THPdb, DrugBank, and PubMed. Use "peptide" and "half-life" as keywords to search PubMed. Use "Plasma" and "Serum" as keywords to search the PEPlife database, collect and organize relevant data. Use "peptide" as a keyword to search the DrugBank and THPdb databases, and collect relevant peptide stability data.

[0032] Step 2: All collected data were cleaned and organized. This included: 1. Excluding peptides with missing or uncertain stability data. 2. Excluding peptides for which corresponding sequence information could not be determined. 3. Excluding peptides with complex modifications that could not be effectively characterized by SMILES codes, such as those modified with polyethylene glycol or bound to macromolecular carriers such as proteins and lipids. Ultimately, 635 data points were successfully screened, covering data from peptide blood stability tests in humans and mice in vitro and in vivo.

[0033] Step 3: After completing data cleaning and organization, SMILES encoding representation was added to all peptides using ChemDraw and RDKit.

[0034] Step 4: Use tools such as AlphaFold2, HighFold, and RDKit to build a three-dimensional structural model for each peptide. For linear peptides containing only natural amino acids and no modifications, AlphaFold2 is used to generate the 3D structure; and for cyclic peptides containing only natural amino acids and no modifications, HighFold is used to generate them. Due to the limitations of current protein prediction tools for peptides containing complex modifications, RDKit (version 2023.3.2) was used to generate the 3D structure of each cyclic peptide. First, 5,000 conformations were generated for each peptide segment, and then the UFF force field was used to optimize the structure of the 5,000 conformations. Finally, the structure with the lowest potential energy was selected as the final conformation of the peptide.

[0035] In step 5, the physicochemical properties of the peptides were calculated using the RDKit Python package. A random forest algorithm was then used to identify the top 140 features that significantly impacted the peptide's blood stability. Species and environmental information were also incorporated into the feature set, resulting in a total of 142 features.

[0036] Step 6: After completing all the previous processing steps, the data is divided into stable and unstable categories based on whether the peptide's half-life in blood is greater than 1 hour. The specific division method is as follows: the data is randomly seeded with 3407 and divided into training and test sets in a ratio of 9:1 for model training and testing.

[0037] Step 7: For the preprocessed data, a multimodal model PepMSND is constructed for training. The 140 molecular descriptors obtained in step 5 and the introduced species and experimental environment information, a total of 142 features, are used as input to the KAN model with a size of N×142 to process the zero-dimensional data of the peptide. The SMILES codes of all peptides are expanded to a length of 1024 and input to the Transformer model with a size of N×1024 to process the one-dimensional data of the peptide. The molecular graph is generated using the DGLpython package and input into GAT to process the two-dimensional data of the peptide. The coordinates are extracted from the PDB file of each peptide and used as input to the SE(3)-transformer to process the three-dimensional data of the peptide.

[0038] Step 8: Resize the outputs of the four modules, KAN, Transformer, GAT, and SE(3)-Transformer, to N×128. Multiply the outputs of these four modules by their corresponding learnable weights W1, W2, W3, and W4. Concatenate the weighted outputs to form a matrix of size N×512. Finally, input this matrix into the subsequent KAN model for dimensionality reduction.

[0039] Step 9: The N×512 matrix concatenated in Step 8 is first passed through a KAN linear layer to reduce its dimensionality to an N×8 matrix. Species and experimental environment information is then incorporated using one-hot encoding and concatenated onto the N×8 matrix, resulting in an N×10 matrix. Finally, this dimensionality reduction process generates a one-dimensional output representing the confidence level of the peptide's stability in blood. The confidence threshold is set to 0.5, with a confidence level greater than 0.5 considered stable and a confidence level less than 0.5 considered unstable.

[0040] The evaluation indicators used in this paper are ACC, Precision, Recall, F1-score, AUC, and MCC. The specific calculation formula is as follows:

[0041]

[0042] Where TP (True Positive), TN (True Negative), FP (False Positive) and FN (False Negative) are used to represent the counts of each category respectively.

[0043] In the embodiments of the present invention, the PepMSND model of the present invention was compared with several existing benchmark machine learning and deep learning methods. The results are shown in Table 1. Compared with these benchmark models, the present invention is in a leading position in all evaluation indicators. In particular, the performance is particularly outstanding in the binary key evaluation indicator AUC, reaching 0.9118. This value fully proves that even when faced with an unbalanced data set, the present invention can still demonstrate an excellent ability to distinguish between positive and negative examples. This strongly illustrates from the side that the present invention has the ability to screen out relatively stable peptide segments from a large number of unstable peptides, and quickly provide lead compounds that meet stability standards for the research and development of polypeptide drugs.

[0044] Table 1. Performance of various models on the peptide blood stability test set

[0045]

[0046]

[0047] The foregoing description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution or modification to the technical solution and technical content disclosed in the present invention, or directly or indirectly applies such changes to other related technical fields, all of which fall within the scope of the present invention and are still within the scope of protection of the present invention.

Claims

1. A deep learning algorithm for predicting the stability of peptides in different blood environments, characterized by The steps include: Step 1: Collect peptide stability data in different blood types from different databases using multiple search keywords; Step 2: Clean and organize the data collected in step 1. Peptides that do not meet the requirements include peptides with missing or uncertain half-life data, peptides with undetermined corresponding sequence information, and peptides with complex modifications that cannot be effectively represented by SMILES codes. Step 3: After data cleaning and organization, SMILES encoding was added to all peptides using ChemDraw 21.0.0 and the Sequence to SMILES.py script; Step 4: After completing step 3, use AlphaFold2, HighFold, and RDKit tools to construct a three-dimensional structure for each peptide; Step 5: Use RDKit to calculate the physicochemical properties of the peptide, and use the random forest algorithm to screen out the top 140 features that have a significant impact on the blood stability of the peptide and retain them; Step 6: After completing all the previous processing steps, the data is divided into stable and unstable categories according to whether the half-life of the peptide in the blood is greater than 1 hour, and the data is divided into training and test sets for model training and testing; Step 7: For the preprocessed data, a multimodal model PepMSND is constructed for training. The model includes a KAN module for receiving zero-dimensional features, a Transformer module for processing one-dimensional features, a GAT module for analyzing two-dimensional features, and a SE(3)-Transformer module for parsing three-dimensional features. Step 8: For the outputs of the four modules KAN, Transformer, GAT, and SE(3)-Transformer of the multimodal model PepMSND, first multiply them by a learnable weight W, and then concatenate the four weighted outputs; Step 9: The spliced ​​output is used as the input of the KAN model, and the characteristic information of the species and experimental environment is added. Finally, a one-dimensional output is generated through dimensionality reduction processing, which represents the confidence level of the peptide stability in blood; Step 10: After model training, the performance of the multimodal model PepMSND is evaluated using the test set. The evaluation indicators include ACC, Precision, Recall, F1-score, AUC, and MCC.

2. A deep learning algorithm for predicting the stability of a polypeptide in different blood environments according to claim 1, characterized in that: The specific operation process of step one is to search and collect relevant data from the public databases of PEPlife, THPdb, DrugBank and PubMed using "peptide", "half-life" and "blood" as keywords.

3. A deep learning algorithm for predicting the stability of a polypeptide in different blood environments according to claim 2, characterized in that: In the second step, a Python-based data cleaning code was used to screen the stability data of peptides that did not meet the requirements in different blood environments based on the collected data.

4. A deep learning algorithm for predicting the stability of a polypeptide in different blood environments according to claim 3, characterized in that: In step 3, the SMILES information of the corresponding peptide was obtained by automatic generation of RDKit and manual drawing of chemdraw.

5. A deep learning algorithm for predicting the stability of a polypeptide in different blood environments according to claim 4, characterized in that: Step 4: For linear peptides containing only natural amino acids and no modifications, AlphaFold2 is used to generate the 3D structure; for cyclic peptides containing only natural amino acids and no modifications, HighFold is used to generate them; for peptides containing complex modifications, RDKit is used to generate the 3D structure of each cyclic peptide.

6. A deep learning algorithm for predicting the stability of a polypeptide in different blood environments according to claim 5, characterized in that: Step 10 uses ACC, Precision, Recall, F1-score, AUC, and MCC to evaluate the performance of the model. The specific calculation formula is as follows: The true positives TP, true negatives TN, false positives FP, and false negatives FN are used to represent the counts of each category respectively.

Images