TWI874059B - A deep learning model system and method thereof - Google Patents

Abstract

A deep learning model system and its method are applied in an environment of predicting and ranking drug target interactions. When using the deep learning system of the present invention to perform a deep learning model method, firstly, a data processing action is performed, and at least one drug data (drug string data (simplified molecular linear input specification (SMILES))) and protein data (protein sequence data (Protein Sequence, for example, one-dimensional protein sequence characteristics)) are processed separately to generate corresponding data respectively. At least one processing result of at least one drug data and protein data. Herein, Word2vec is used to learn the at least one protein sequence data (Protein Sequence, for example, one-dimensional protein sequence characteristics) and the SMILES string (at least one drug String data). Word2vec is one of the deep learning natural language processing (NLP) methods, and the natural language methods is used to explore and deepen the understanding of life languages. It follows the grammar encoded in protein amino acid sequences and SMILES strings to learn and obtain low-dimensional feature vectors of these sequences and strings. Then, Perform a training action and import the processing results into a predictive affinity model for training. Herein, these representations are combined with protein docking interaction features and put into a deep model for prediction; then, the action of generating a prediction result is performed, and a prediction result is generated through the training of the prediction affinity model.

Description

A deep learning model system and method thereof

The present invention relates to a deep learning model system and a method thereof, and more specifically, to a deep learning model system and a method thereof applied in the environment of drug-target interaction prediction and ranking, and belongs to the technical field of drug-target interaction prediction and ranking using deep learning.

It often takes a lot of time and money to develop a drug and bring it to market. For new drug research and development to registration and marketing, it takes an average of 13 to 15 years, and the cost is an average of 2 to 3 billion. According to statistics, these costs are still on the rise. In order to reduce the cost of money and time, the pharmaceutical industry is currently replacing some new drug development with drug reorganization. The spirit of drug reorganization comes from the cost consideration and analysis of economic pros and cons, and accelerates the drug development process by confirming new clinical uses for existing drugs. The success and increasing application of drug reorganization can be regarded as one of the extensions of multi-target pharmacology, which represents the transformation from a single-target form to a multi-target form in drug discovery. However, whether in drug discovery or drug reorganization, identifying the interaction between drugs and targets (DTI) is one of the important steps. Therefore, identifying DTI has become a crucial topic in research in related fields. However, current identification methods still rely on web lab experiments to observe and record the interaction between drugs and target molecules (such as enzymes, cell surface receptors, G protein-coupled receptors (GPCRs), ligand-gated ion channels, nuclear receptors, etc.) to prove their efficacy. The resulting time and financial costs are still major obstacles in many drug research.

With the rapid development of intelligent computing technology, the reuse of existing drugs and the exploration of new interactions through computational simulation have begun to attract attention. In earlier studies, researchers used docking simulation software such as Autodock Vina, GOLD, ZDOCK, RDOCK, and HADDOCK to predict and sort the docking sites between drugs and targets. In addition to predicting and sorting the docking sites, docking simulation software can also predict binding affinity. However, docking simulation software simplifies many calculation formulas due to the consideration of computational time cost, so that many important information for affinity is not considered, such as factors such as water solvent or hydrogen bond strength, resulting in poor performance of docking simulation software in predicting binding affinity. Although it is possible to combine molecular dynamics simulation software, such as Molecular Dynamics simulation (MD simulation) or Quantum-Mechanics (QM), to simulate the dynamic information of proteins and analyze binding sites, and to simulate the drug binding stage and provide a deep understanding of DTI affinity, and quantify the energy changes in the entire process, thereby improving the prediction ability, this must take a lot of time and cost. In addition to simulation software, some researchers have explored the use of a large amount of drug-target interaction information in the database and combined it with computer science, and tried to improve the existing computing costs and improve the prediction effect. There are still many unpaired small molecule compounds to be explored in drug-target interactions and there is the possibility of developing new drugs. Currently, many protein experimental databases (such as BindingDB, PCI-DB, etc.) already have many known interactions, but these known drug targets have rarely been verified. Although compound databases (such as ChEMBL, PubChem) have a large amount of compound data to provide drug information, there are still many unknown interactions waiting to be discovered. In terms of recent developments, DTI recognition has expanded from early binary classification to regression problems. Many early studies focused on binary classification. For example, Bleakley and Yamanishi used bipartite local models (BLMs) to predict the interactions between four types of protein targets commonly used in pharmacology (enzymes, GPCRs, ion channels, and nuclear receptors) and drugs in humans, and provided data sets of these four types for future researchers to use in studying interactions. van Laarhoven et al. used the Gaussian Interaction Profile kernel to predict the binary classification of drug-target interactions. Hakime Öztürk et al. used the Weighted Nearest Neighbor-Gaussian Interaction Profile (WNN-GIP) to predict the binary classification of drug-target interactions based on the one-dimensional drug data type (simplified molecular input line entry specification, SMILES) similarity function, find the similarity between compounds, and other related research. More recently, some research has expanded to explore regression problems. Compared with binary classification, which requires self-determining thresholds to classify drug-target pairs into those with or without inhibitory effects, there is no information on strong or weak affinity concentrations, and the binding affinity is directly predicted. Affinity) can provide researchers with accurate inhibition concentrations for subsequent experiments. Binding affinity provides information about the strength of the interaction between drug-target (DT) pairs. Depending on different formats and mathematical formulas, affinity is expressed as dissociation constant (Kd), inhibition constant (Ki) or half of the maximum inhibitory concentration (IC50). IC50 depends on the concentration of the target and ligand, and a low concentration IC50 indicates a strong binding affinity. Similarly, a low Ki value indicates a strong binding affinity. Kd and Ki values are usually expressed as pKd or pKi, which is the negative logarithm of the dissociation constant or inhibition constant. T Pahikkala et al. used a regression model called Kronecker regularized least-squares method (Kronecker regularized least-squares method, KronRLS) explored four different training and validation set cutting methods for real-world drug-target pairing examples, and predicted the drug-target interaction affinity Kd and Ki constants; Hakime Öztürk et al. used a convolutional neural network called DeepDTA, and used it for protein and drug sequence character data to predict and sort the affinity Kd and Ki constants of drug-protein interactions, and improved DeepDTA in 2019, adding protein and drug sequence data of different forms, called WIDEDTA; Jiménez, J. et al. used a convolutional neural network (CNN) for drug-target docking 3D structural data to predict the junction of drug-target interactions; Mostafa In 2019, Karimi et al. combined CNN and deep neural network (DNN) to use protein sequences in the PDBbindDB database and drug SMILES sequences in the STITCH database to predict drug-target interaction affinity IC50, Kd, and Ki constants.

In the studies conducted so far, many studies have used different preprocessing methods based on their own considerations (such as protein sequence preprocessing, drug string preprocessing, etc.), and few studies have considered structural information. In addition, many studies are based on the four protein datasets commonly used in pharmacology provided by Yamanishi et al.

The technical features of the paper "AttentionDTA: prediction of drug–target binding affinity using attention model" (author Qichang Zhao) in the non-patent document "2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)" are: With the advancement of biotechnology and the discovery and expansion of new drugs in the pharmacology community, the reuse of existing drugs and the exploration of new interactions have also attracted the attention of researchers. Compared with traditional modeling-based methods (such as docking or molecular dynamics simulation), big data is now on the rise, and the prediction of computer interactions is also driven by big data. With recent developments, DTI recognition has expanded from early binary classification to regression problems. Many early studies focused on binary classification, such as Bleakley and Yamanishi, who used bipartite local models (BLMs) in 2009 to predict the binary classification of four important types of drug-target interactions in humans, and provided these four types of data sets for future interaction researchers; van Laarhoven et al., who used the Gaussian Interaction Profile kernel in 2011 to predict the binary classification of drug-target interactions. In 2016, Hakime Öztürk et al. used the weighted nearest neighbor-Gaussian interaction profile (WNN-GIP) to find the similarity between compounds based on the SMILES similarity function of the one-dimensional drug data type. Recently, with the rise of data, many studies have expanded to explore regression problems. For example, T Pahikkala et al. used a regression model called Kronecker regularized least-squares method (KronRLS) in 2014 to classify drug targets into four types for real-world application examples, and predicted and ranked drug-target interaction affinity; Hakime Öztürk et al. used convolutional neural network (Convolutional neural In 2018, Jiménez, J. et al. used convolutional neural network (CNN) to dock 3D structure data of drugs and targets to predict the junction of drug-target interactions; Mostafa Karimi et al., in 2019, combined CNN and RNN to predict the affinity of drug-target interactions using protein sequences in the PDBindDB database and drug SMILES sequences in the STITCH database. Binding affinity provides information about the strength of the interaction between a drug-target (DT) pair, and is expressed in terms of the dissociation constant (Kd), inhibition constant (Ki), or half maximal inhibitory concentration (IC50) depending on the format and mathematical formula used. The IC50 depends on the concentration of the target and ligand, and low concentrations of IC50 indicate strong binding affinity. Similarly, low Ki values indicate strong binding affinity. Kd and Ki values are usually expressed as pKd or pKi, which are the negative logarithms of the dissociation constant or inhibition constant. Many studies to date have been related to, referenced, or used the four datasets provided by the Yamanishi et al. study, for example, the non-patent reference. Most importantly, the non-patent document does not disclose the technical features of Word2vec, let alone using Word2vec (a related model used to generate word vectors (word2vec)) to learn one-dimensional protein sequences and ligand SMILES strings. Word2vec is applied to deep learning natural language processing (NLP) methods.

So, how can we avoid using the following shortcomings of the known techniques: Many studies focus on binary classification, such as Bleakley and Yamanishi using a bipartite local model to predict the interactions between four types of protein targets commonly used in medicine (enzymes, GPCRs, ion channels, nuclear receptors) and drugs in humans, and provide these four types of data sets for future researchers to use in studying interactions; van Laarhoven et al. used Gaussian interaction spectrum functions to predict the binary classification of drug-target interactions; Hakime Öztürk et al. used weighted nearest neighbor Gaussian interaction spectrum based on the one-dimensional data type of drugs (simplified molecular input line entry specification, SMILES) similarity function, find the similarity between compounds and other related research; more recently, some research has begun to expand to explore regression problems. Compared with binary classification, which requires self-determining thresholds to classify drug-target pairs into those with or without inhibitory effects, there is no strong or weak affinity concentration information, and direct prediction of binding affinity is a problem to be solved.

The main purpose of the present invention is to provide a deep learning model system and method thereof, which are applied in the environment of drug-target interaction prediction and ranking. When the deep learning system of the present invention is used to perform the deep learning model method, first, a data processing operation is performed to perform a data processing on at least one drug data (drug string data (ligand simplified molecular linear input specification (SMILES)) and protein data (protein sequence data (Protein Sequence, for example, one-dimensional protein sequence features)) respectively, and at least one processing result corresponding to the at least one drug data and protein data is generated. Here, Word2vec is used to learn the at least one protein sequence data (Protein Sequence, for example, one-dimensional protein sequence features) and the ligand SMILES. String (at least one drug string data (ligand simplified molecular linear input specification (SMILES)), Word2vec is applied to one of the deep learning natural language processing (NLP) methods, and uses natural language methods to explore and deepen the understanding of life language, following the grammar encoded in the protein amino acid sequence and the SMILES string to learn and obtain low-dimensional feature vectors of these sequences and strings; then, a training action is performed, and the processing results are introduced into a prediction affinity model for training, where these representations are combined with protein ligand docking interaction features and put into the deep model for prediction; then, a prediction result generation action is performed, and a prediction result is generated through the training of the prediction affinity model; and the prediction results are sorted.

Another object of the present invention is to provide a deep learning model system and method thereof, which is applied in the context of drug-target interaction prediction and ranking, discloses the technical features of Word2vec, and uses Word2vec (a related model for generating word vectors (word2vec)) to learn one-dimensional protein sequences and ligand SMILES strings. Applied to deep learning natural language processing (NLP) methods; in the present invention, in addition to the amino acid characteristics and chemical formula information in the primary structure (sequence) of the protein and the drug string, the three-dimensional structure of the protein and the drug also has extremely important information that allows the model to learn more comprehensive drug-target docking features. Therefore, the protein (target) 3D structure and the automatically generated binding pattern of the drug (ligand) in the docking simulation software can be used to calculate and obtain the structure-based interaction characteristics between the protein and the drug. These features are combined with other data types, such as one-dimensional protein sequence data and drug SMILES strings, and a method for combining these data types to predict and establish the binding affinity of protein-drug interactions in deep learning models is proposed. Method, for this purpose, the present invention utilizes a correlation model (word2vec) for generating word vectors to learn one-dimensional protein sequences and ligand SMILES strings. The correlation model for generating word vectors is applied to one of the deep learning natural language processing (NLP) methods. Through existing natural language methods, the understanding of life language is explored and deepened. The Word2vec model, one of the deep learning natural language processing (NLP) methods, is applied to learn one-dimensional protein sequences and drug SMILES strings. The grammar encoded in the protein amino acid sequence and the SMILES string is followed to learn and obtain the mapping of these sequences and strings to feature vectors, and the converted vectors are combined with interaction features and input into the deep model for prediction. In addition, it is proposed to use Multi-Modal to learn these data representations and interaction characteristics and predict binding affinity, and compare with common machine learning models such as Random Forest and Gradient Boost. The results show that the performance of combining different types of data, whether it is Mean Square Error (MSE) or Concordance Index (C-index), is improved by 10~15% compared with the results of using only one type of data. In the future, we hope to predict more accurate and better ranked drug-target interactions by combining more types of data, and establish a platform to perform multiple analyses on many protein families and polypharmacology to explore unknown characteristics.

Another purpose of the present invention is to provide a deep learning model system and method thereof, which are applied in the environment of drug-target interaction prediction and ranking. The present invention does not use the data set used by the above-mentioned known technology. Considering factors such as multiple and superfamily analysis, the drug target data set studied by Daniel A Bachovchin et al. is adopted. The target used in the technical feature of the present invention is to explore the inhibitory effect of enzyme activity. The enzyme is serine hydrolase. Serine hydrolase is one of the largest and most diverse enzyme species known in organisms. In mammals, it accounts for about 1% of all proteins, and in the human kingdom, it accounts for about 240 members. Serine hydrolases play many key roles in biological processes, such as blood homeostasis, neural signaling and glucose homeostasis in pathological and physiological processes. They also have important functions in viruses and bacteria, such as bacterial and viral infections, life cycles, toxicity and drug resistance. Among the more than 200 physiological compounds and functions of human serine hydrolases, many do not have good characteristics and most lack selective active inhibitors. We try to perform high-throughput and multiplex analysis of serine hydrolases, explore their characteristics, and make an analysis platform. We use 96 serine hydrolase members and 52 drugs for experimental observation and record the inhibition concentration and the characteristics after interaction with each other, so as to improve the selectivity of inhibitors for more human serine hydrolases. The present invention is different from the above-mentioned prior art and literature. Considering that traditional docking simulation software such as Autodock Vina, GOLD, ZDOCK, RDOCK, and HADDOCK can predict binding affinity in addition to predicting and sorting docking points, docking simulation software simplifies many calculation formulas due to the consideration of calculation time cost, resulting in many important information for affinity not being considered, such as factors such as water solvent or hydrogen bond strength, resulting in poor performance of docking simulation software in predicting binding affinity. Although it can be combined with molecular dynamics simulation (MD simulation) or Quantum-Mechanics (QM) to increase accuracy, this must take a lot of time and cost. In the present invention, the protein 3D structure and the automatically generated binding pattern of the ligand in the docking simulation software are used to calculate and obtain the structure-based interaction characteristics between the protein and the ligand. These characteristics are combined with other data types, such as one-dimensional protein sequence data and ligand SMILES strings, and a method for combining these data types to predict the binding affinity of protein-ligand interaction with a deep learning model is proposed. For this purpose, the applicant uses Word2vec (a related model used to generate word vectors (word2vec)) to learn one-dimensional protein sequences and ligand SMILES strings. Word2vec is applied to one of the deep learning natural language processing (NLP) methods, and uses existing natural language methods to explore and deepen the understanding of life language. According to the protein amino acid sequence and SMILES The grammar encoded in the string is used to learn and obtain the low-dimensional feature vectors of these sequences and strings, and these representations are combined with the interaction features and put into the deep model for prediction. In addition, it is also proposed to use Multi-Modal to learn these data representations and interaction features and predict binding affinity, and compare them with common machine learning models such as Random Forest and Gradient Boost. The results show that the performance of combining different types of data, whether it is the Mean Square Error (MSE) or the Concordance Index (C-index) ranking, is improved by 10~15% compared with the results of using only one type of data. In the future, we hope to combine more types of data to predict more accurate and better-ranked drug-target interactions, and establish a platform to perform multiple analyses on many protein families and polypharmacology to explore unknown properties.

According to the above-mentioned purpose, the present invention provides a deep learning model system, which includes a data processing service module, a predicted affinity model module, and a ranking model module.

A data processing service module, wherein the data processing service module processes at least one drug data (drug string data (SMILES)) and protein data (protein sequence data (Protein Sequence, for example, one-dimensional protein sequence features)) respectively, and generates at least one processing result corresponding to the at least one drug data and protein data respectively; the data processing service module processes at least one drug data and protein data respectively, and generates at least one processing result corresponding to the at least one drug data and protein data respectively, wherein the drug data and protein data here include the simplified molecular input line entry specification (SMILES) of the drug. SMILES) D_drug, protein sequence D_protein and drug protein structure data D_drug_protein, the data processing can be generated by a data processing device that executes a natural language processing (word2vec) model, and the processing result is at least one sequence data, and is integrated with the drug protein structure data D_drug_protein.

Here, Word2vec is used to learn the at least one protein sequence data feature (Protein Sequence, for example, one-dimensional protein sequence feature) and the drug string data SMILES (ligand SMILES string (at least one drug string data (ligand simplified molecular linear input specification (SMILES))). Word2vec is applied to one of the deep learning natural language processing (NLP) methods, and uses natural language methods to explore and deepen the understanding of life language. The low-dimensional feature vectors of these protein sequences and drug strings are learned and obtained according to the grammar encoded in the protein sequence data feature amino acid sequence and the drug string data SMILES.

Each data processed by the data processing service module includes three types: protein-ligand docking interaction features, protein sequences, and ligand simplified molecular linear input specifications SMILES (i.e., drug data (drug string data)), which are obtained from different platforms; protein-ligand docking interaction features are extracted by simulation by Autodock Vina software, for example, 39 protein-ligand docking interaction feature names. Protein sequences are ProteinIDs in the data that are compared and obtained using the UniProt database. UniProt is a comprehensive database of protein sequences and functions, many of which come from genome projects and contain a large amount of information on the biological functions of proteins from research literature and many studies. The SMILES uses the CID in the data to compare with the PubChem database and obtain the sequence. The PubChem database provides biological activity data of small organic molecules related to biological activities. It is maintained by the National Center for Biotechnology Information (NCBI), a subsidiary of the National Institutes of Health (NIH). PubChem is part of the NIH "Molecular Libraries Roadmap Initiative". Information can be accessed for free at www. You can also use FTP to download millions of chemical databases for free. The PubChem database can query various organic chemical molecules, including drugs and chemical raw materials.

The drug data includes a ligand (drug) simplified molecular linear input specification SMILES D_drug, and the protein sequence data (protein data) includes a protein sequence D_protein, and the protein-ligand docking interaction feature includes drug protein structure data D_drug_protein. The data processing of the data processing service module can be generated by executing a natural language processing word2vec model to generate the processing result, and the processing result is at least one sequence data, and is integrated with the drug protein structure data D_drug_protein of the protein-ligand docking interaction feature.

The data processing service module uses the technical features of Word2vec and uses Word2vec (a related model used to generate word vectors (word2vec)) to learn the one-dimensional protein sequence and ligand SMILES string. Applied to deep learning natural language processing (NLP) methods; in the present invention, in addition to the amino acid characteristics and chemical formula information in the primary structure (sequence) of the protein and the drug string, the three-dimensional structure of the protein and the drug also has extremely important information that allows the model to learn more comprehensive drug-target docking features. Therefore, the protein (target) 3D structure and the automatically generated binding pattern of the drug (ligand) in the docking simulation software can be used to calculate and obtain the structure-based interaction characteristics between the protein and the drug. These features are combined with other data types, such as one-dimensional protein sequence data and drug SMILES strings, and a method for combining these data types to predict and establish the binding affinity of protein-drug interactions for deep learning models is proposed. This invention uses a related model (word2vec) for generating word vectors to learn one-dimensional protein sequences and ligand SMILES strings. The related model for generating word vectors is applied to one of the deep learning natural language processing (NLP) methods. The existing natural language methods are used to explore and deepen the understanding of life language. The Word2vec model, one of the deep learning natural language processing (NLP) methods, is applied to learn one-dimensional protein sequences and drug SMILES strings. The grammar encoded in the protein amino acid sequence and the SMILES string is followed to learn and obtain these sequences and strings mapped to feature vectors, and the converted vectors are combined with interaction features and input into the prediction affinity model module for prediction.

The present invention is different from the above-mentioned prior art and literature. Considering that traditional docking simulation software such as Autodock Vina, GOLD, ZDOCK, RDOCK, and HADDOCK can predict binding affinity in addition to predicting and sorting docking locations, docking simulation software simplifies many calculation formulas due to the consideration of calculation time cost, resulting in many important information for affinity not being considered, such as factors such as water solvent or hydrogen bond strength, resulting in the docking simulation software not having a very good performance in predicting binding affinity. Although it can be combined with molecular dynamics simulation (MD simulation) or Quantum-Mechanics (QM) to increase the accuracy, this must take a lot of time and cost.

In the present invention, the protein 3D structure and the automatically generated binding pattern of the ligand in the docking simulation software are used to calculate and obtain the structure-based interaction characteristics between the protein and the ligand. These characteristics are combined with other data types, such as one-dimensional protein sequence data and ligand SMILES strings, and a method for combining these data types to predict the binding affinity of protein-ligand interaction with a deep learning model is proposed. For this purpose, Word2vec (a related model used to generate word vectors (word2vec)) is used to learn the one-dimensional protein sequence and ligand SMILES string. Word2vec is applied to one of the deep learning natural language processing (NLP) methods, and the existing natural language methods are used to explore and deepen the understanding of the language of life. According to the protein amino acid sequence and SMILES The grammar encoded in the strings is learned to obtain low-dimensional feature vectors of these sequences and strings, and these representations are combined with interaction features and then put into the deep model for prediction.

A prediction affinity model module, wherein the data processing service module utilizes the technical features of Word2vec, and uses Word2vec (a related model (word2vec) for generating word vectors) to learn protein sequence data, for example, a one-dimensional protein sequence and a string of a ligand (drug) simplified molecular linear input specification SMILES, and Word2vec is applied to a deep learning natural language processing (NLP) method; in the present invention, in addition to the amino acid characteristics and chemical formula information in the primary structure (sequence) of the protein and the drug string, the three-dimensional structure of the protein and the drug also has extremely important information that allows the prediction affinity model module to learn more comprehensive drug-target docking features (protein-ligand docking interaction features).

Therefore, by using the data processing service module and/or the affinity prediction model module, the protein (target) 3D structure and the automatically generated binding pattern of the drug (ligand) in the docking simulation software can be used to calculate and obtain the structure-based interaction characteristics (protein-ligand docking interaction characteristics) between the protein and the drug, and these characteristics (protein-ligand docking interaction characteristics) are combined with other data types, such as one-dimensional protein sequence data and drug SMILES strings, and a method for combining these data types to predict and establish a deep learning model for protein-drug interaction binding affinity is proposed. For this purpose, the present invention uses the Word2 related model used to generate word vectors vec is used to learn one-dimensional protein sequences and ligand SMILES strings. The related model used to generate word vectors is applied to one of the deep learning natural language processing (NLP) methods. Through existing natural language methods, we explore and deepen the understanding of life language. The Word2vec model, one of the deep learning natural language processing (NLP) methods, is applied to learn one-dimensional protein sequences and drug SMILES strings. The grammar encoded in the protein amino acid sequence and SMILES string is learned and mapped to feature vectors. The converted vectors are combined with interaction features and input into the prediction affinity model module for prediction.

In addition, in the affinity prediction model module, it is also proposed to use Multi-Modal to learn these data representations and interaction characteristics and predict binding affinity, and compare them with common machine learning models such as Random Forest and Gradient Boost. The results show that the performance of combining different types of data, whether it is the Mean Square Error (MSE) or the Concordance Index (C-index), is improved by 10~15% compared with the results of using only one type of data. In the future, we hope to predict more accurate and better ranked drug-target interactions by combining more types of data, and establish a platform to perform multiple analyses on many protein families and polypharmacology to explore unknown characteristics.

The target used in the technical features of the present invention is to explore the inhibitory effect of enzyme activity, and the enzyme is serine hydrolase. Serine hydrolase is one of the largest and most diverse enzyme species known in organisms. In mammals, it accounts for about 1% of all proteins, and in humans, there are about 240 members. Serine hydrolases play many key roles in biological processes, such as blood homeostasis, neural signaling and glucose homeostasis in pathological and physiological processes. They also have important functions in viruses and bacteria, such as bacterial and viral infections, life cycles, toxicity and drug resistance. Among the more than 200 physiological compounds and functions of human serine hydrolases, many do not have good characteristics and most lack selective active inhibitors. We try to conduct high-throughput, multiplex analysis of serine hydrolases, explore their characteristics, and create an analysis platform. We use 96 serine hydrolase members and 52 drugs to conduct experimental observations and record the inhibition concentration and the characteristics after interaction with each other, so as to improve the selectivity of inhibitors for more human serine hydrolases.

The prediction affinity model module imports at least one of the processing results into the prediction affinity model module for training, so as to generate a prediction result through the training of the prediction affinity model of the prediction affinity model module, that is, the sequence data and the structure data are imported into the prediction affinity model module which can be a convolutional neural model/deep neural model for training to generate a prediction result, that is, the drug vector and the protein vector generated by the above data processing are processed to generate the prediction result, and the prediction result generated is integrated with the drug protein structure data D_drug_protein by using the data processing service module and/or the prediction affinity model module.

A sorting model module is used to sort the prediction results; the prediction results are integrated with the drug protein structure data D_drug_protein and the results are introduced into the sub-network of the deep network model. Here, for example, the predicted affinity model module is a deep network model, and the generated multi-modal architecture model is then introduced into the predicted affinity model module formed by the sub-network to be processed to generate the mean square error D_MSE, and the predicted affinity model module transmits the D_MSE to the sorting model module; then, the sorting is performed through the sorting model module to obtain the consistency index D_CI.

Here, in one embodiment, the data processing is to process the drug data and protein data according to their data types, and the processing results generated are at least one sequence data and at least one structure data, and the predicted affinity model module is a deep network model.

The data processing service module uses the technical features of Word2vec and uses Word2vec (a related model (word2vec) used to generate word vectors) to learn one-dimensional protein sequences and ligand SMILES strings. Word2vec is applied to deep learning natural language processing (NLP) methods. In the present invention, in addition to the amino acid characteristics and chemical formula information in the primary structure (sequence) of the protein and the drug string, the three-dimensional structure of the protein and the drug also has extremely important information that allows the model to learn more comprehensive drug-target docking features. Therefore, the protein (target) 3D structure and the automatically generated binding mode of the drug (ligand) in the docking simulation software can be used to calculate and obtain the structure-based interaction characteristics between the protein and the drug.

In the present invention, the protein 3D structure and the automatically generated binding pattern of the ligand in the docking simulation software are used to calculate and obtain the structure-based interaction characteristics between the protein and the ligand. These characteristics are combined with other data types, such as one-dimensional protein sequence data and ligand SMILES strings, and a method for combining these data types to predict the binding affinity of protein-ligand interaction with a deep learning model is proposed. For this purpose, Word2vec (a related model used to generate word vectors (word2vec)) is used to learn the one-dimensional protein sequence and ligand SMILES string. Word2vec is applied to one of the deep learning natural language processing (NLP) methods, and the existing natural language methods are used to explore and deepen the understanding of the language of life. According to the protein amino acid sequence and SMILES The grammar encoded in the strings is learned to obtain low-dimensional feature vectors of these sequences and strings.

In order to make those familiar with the art understand the purpose, features and effects of the present invention, the present invention is described in detail by the following specific embodiments and the accompanying drawings:

Protein-ligand Pair Docking is a method in the field of molecular modeling that predicts the orientation of one molecule relative to another when a protein binds to a target to form a stable complex. It is a commonly used method for structural drug design, and the orientation results can be used to predict the affinity strength between two molecules through scoring functions. Protein-ligand docking technology has long been used to predict new drugs and drug recombination. Common simulation software includes Autodock, ZDOCK, RDOCK, and HADDOCK. However, docking will generate multiple different positions for each ligand and further compare and rank them with different compounds based on the scoring function. However, the simplified calculation formula may lead to inaccurate predictions of docking affinity, despite accurate predictions of docking positions, so this is not the best tool for predicting affinity. The present invention collects the features obtained by the docking evaluation method used in Autodock Vina as data features for the deep model prediction of the present invention. We have a total of 39 docking features as inputs to the deep model.

The amino acid sequence (or residue sequence) of a peptide or protein is the primary structure of the protein. It is the linear sequence of amino acids in the peptide or protein. The amino acids in the primary structure can interact with each other and form hydrogen bonds with amine groups using the C=O bond on the amide bond. Based on the nature of the free groups at each end, the two ends of the polypeptide chain are called the carboxyl terminus (C-terminus) and the amino terminus (N-terminus). The counting of residues always starts at the N terminus (NH 2-group), which is the end where the amine group does not participate in the peptide bond. The primary structure of the protein is determined by the gene corresponding to the protein. The specific nucleotide sequence in the DNA is transcribed into mRNA, and the ribosome reads the sequence in a process called translation. The amino acid sequence in insulin was discovered by Frederick Sanger, establishing that proteins have a definite amino acid sequence. The sequence of a protein is unique to that protein and defines the structure and function of the protein. The sequence of a protein can be determined by methods such as Edman degradation or tandem mass spectrometry. However, it is more often read directly from the gene sequence using the genetic code. When discussing proteins, it is strongly recommended to use the term "amino acid residues" because when a peptide bond is formed, a water molecule is lost, so the protein is composed of amino acid residues. Post-translational modifications such as phosphorylation and glycosylation are also usually considered part of the primary structure and cannot be read from the gene. For example, insulin is composed of 51 amino acids in 2 chains. One chain has 31 amino acids and the other has 20 amino acids.

Protein sequences are usually represented by a string of letters, listing the amino acids from the amino terminus to the carboxyl terminus, with the letters representing the amino acid, represented by a three-letter code or a single letter code, with the letters representing the 20 naturally occurring amino acids, as well as mixtures or unusual amino acids (similar to nucleic acid symbols).

The Simplified Molecular Input Line Entry Specification (SMILES) is a simple yet comprehensive chemical language, a line notation specification that can be used to unambiguously describe the structure of a chemical substance or specify molecules and reactions using ASCII strings representing atoms and bond symbols. SMILES contains the same information as the extended table of elements, but has several advantages over it. In practice, any SMILES string representing a molecule or reaction can be used to give a chemically correct and understandable description, and if the specification represents a unique string, it can be used as a universal identifier string for a specific chemical structure. Most molecular editors can import SMILES strings to convert back into two-dimensional graphics and three-dimensional models of the molecule. The original SMILES specification, originated in the 1980s by David Weininger, used the concept of a graph with nodes as atoms, edges as bonds to represent molecules, brackets to indicate branch points, and numeric labels to indicate ring junctions. The basic SMILES syntax also includes isotopic information, regarding the configuration and chirality of double bonds; these syntaxes are called isomeric SMILES. Since its development, SMILES has been modified and extended by many, notably Daylight Chemical Information Systems Inc., to include not only new functionality but also two other chemical languages: SMARTS®, an extension of SMILES that allows molecular patterns and properties to be specified at varying levels of specificity for use in substructure searches; and SMIRKS®, a restricted version of reaction SMARTS® involving variations of atomic bonding patterns to define general reactions. In 2007, the Open Source Chemistry Group developed an open standard called OpenSMILES. Other linear notations include the Wiswesser Line Notation (WLN), ROSDAL, and the SYBYL Line Notation (SLN).

FIG. 1 is a system schematic diagram for illustrating the system architecture and operation of the deep learning model system of the present invention. As shown in FIG. 1 , the deep learning model system 1 includes a data processing service module 2, a predicted affinity model module 3, and a ranking model module 4.

The data processing service module 2, the predicted affinity model module 3, and the ranking model module 4 are at least one of hardware, firmware, and software, and operate in conjunction with a processor (not shown) of an electronic device where the deep learning model system 1 is located. Here, the electronic device may be, for example, a personal computer, a server, an AI server, an Android phone, an iPhone, a tablet computer, an iPad, etc.

A data processing service module 2 performs data processing on at least one drug data and protein data respectively, and generates at least one processing result corresponding to the at least one drug data and protein data respectively. The drug data and protein data here include a simplified molecular input line entry specification (SMILES) D_drug, a protein sequence D_protein and a drug protein structure data D_drug_protein. The data processing can be performed by the data processing service module 2 executing a natural language processing (word2vec) model to generate the processing result, and the processing result is at least one sequence data, and is integrated with the drug protein structure data D_drug_protein.

Predictive affinity model module 3, data processing service module 2 imports the processing result into the predictive affinity model module 3 for training, so as to generate a prediction result through the training of the predictive affinity model module 3, wherein the processing result is at least one sequence data and at least one structural data, that is, the sequence data and the structural data are imported into the predictive affinity model module 3 which can be a convolutional neural model/deep neural model for training to generate a prediction result, that is, the drug vector and protein vector generated by the above data processing are processed to generate a prediction result, and integrated with the drug protein structure data D_drug_protein.

The prediction results are sorted, and the results of integrating the prediction results with the drug protein structure data D_drug_protein are introduced into the sub-network of the deep network model. The generated multi-modal architecture model is then introduced into the prediction affinity model module 3 formed by the sub-network for processing to generate the mean square error D_MSE.

The sorting model module 4 will receive the mean square error D_MSE from the predicted affinity model module 3, and sort through the sorting model module 4 to obtain the consistency index D_CI; the data processing is to process the drug data and protein data according to their data types, and the processing results generated are at least one sequence data and at least one structure data, and the predicted affinity model module 3 is a deep network model.

In the present invention, in addition to protein sequences and drug strings, structural information of protein-drug interactions is considered, and these three types of information are applied to the model, and pre-processed according to the rules of protein property and drug string conversion. In addition, many studies are based on four protein data sets commonly used in pharmacology provided by Yamanishi et al., while in the present invention, factors such as multiple and superfamily analysis are considered, and the data sets used by the above authors are not used, but the drug target data set studied by Daniel A Bachovchin et al. is used. The target used in this study is to explore the inhibitory effect of serine hydrolase activity. Serine hydrolase is one of the most numerous and diverse enzyme types known in organisms. In mammals, it accounts for about 1% of all proteins, and in humans, it accounts for about 240 protein members. Serine hydrolases play many key roles in biological processes, such as blood homeostasis, neural signaling and glucose homeostasis in pathophysiological processes. They also have important functions in viruses and bacteria, such as bacterial and viral infections, life cycles, toxicity and drug resistance. In previous studies, researchers have proposed more than 200 physiological compounds in human serine hydrolases, and tried to perform high-throughput and multiplex analysis of serine hydrolases to explore their characteristics and create an analytical platform, which used 96 serine hydrolase members and 52 drugs for experimental observation and recorded the inhibition concentration and the characteristics after interaction with each other, in order to improve the selectivity of inhibitors for more human serine hydrolases.

In the present invention, in addition to the amino acid characteristics and chemical formula information in the primary structure (sequence) of the protein and the drug string, the three-dimensional structure of the protein and the drug also has extremely important information that allows the model to learn more comprehensive drug-target pair characteristics. Therefore, the protein (target) 3D structure and the automatically generated binding pattern of the drug (ligand) in the traditional docking simulation software are used to calculate and obtain the structure-based interaction characteristics between the protein and the drug. These characteristics are combined with other data types, such as one-dimensional protein sequence data and drug SMILES strings, and a method for combining these data types to predict and establish the binding affinity of protein-drug interactions in a deep learning model is proposed. For this purpose, the present invention uses The Word2vec model used to generate word vectors is used to learn one-dimensional protein sequences and ligand SMILES strings. The Word2vec model used to generate word vectors is applied to one of the deep learning natural language processing (NLP) methods. Through existing natural language methods, we explore and deepen our understanding of life languages. We apply the Word2vec model, one of the deep learning natural language processing (NLP) methods, to learn one-dimensional protein sequences and drug SMILES strings. We follow the grammar encoded in the protein amino acid sequence and SMILES string to learn and map these sequences and strings to feature vectors, and combine the converted vectors with the interaction features and input them into the deep model for prediction.

In addition, it is proposed to use Multi-Modal to learn these data representations and interaction characteristics and predict binding affinity, and compare them with common machine learning models such as Random Forest and Gradient Boost. The results show that combining different types of data improves the mean square error (MSE) or consistency index (C-index) ranking by 10~15% compared with the results of using only one type of data. In the future, we hope to predict more accurate and better ranked drug-target interactions by combining more types of data, and establish a platform to perform multiple analyses on many protein families and polypharmacology to explore unknown characteristics.

FIG. 2 is a flow chart illustrating the process steps of performing a deep learning model method using the deep learning model system of the present invention as shown in FIG. 1 .

In step 101, data processing is performed; the data processing service module 2 performs data processing on at least one drug data (drug string data (ligand simplified molecular linear input specification (SMILES)) and protein data (protein sequence data (Protein Sequence, for example, one-dimensional protein sequence characteristics)) to generate at least one processing result corresponding to the at least one drug data and protein data, respectively. Here, Word2vec is used to learn the at least one protein sequence data (Protein Sequence, for example, one-dimensional protein sequence characteristics) and the ligand SMILES string (at least one drug string data (ligand simplified molecular linear input specification (SMILES)), Word2vec One of the deep learning natural language processing (NLP) methods is applied, and a natural language method is used to explore and deepen the understanding of the language of life. The low-dimensional feature vectors of these sequences and strings are learned and obtained according to the grammar encoded in the protein amino acid sequence and the SMILES string, and then the step 102 is entered.

Here, in step 101, the data processing service module 2 performs data processing on at least one drug data and protein data respectively, and generates at least one processing result corresponding to the at least one drug data and protein data respectively. The drug data and protein data here include the simplified molecular input line entry specification (SMILES) D_drug, the protein sequence D_protein and the drug protein structure data D_drug_protein. The data processing can be performed by the data processing service module 2 executing the natural language processing (word2vec) model to generate the processing result, and the processing result is at least one sequence data, and is integrated with the drug protein structure data D_drug_protein.

As for Word2Vec, the present invention uses the skip-gram model, so only Skip-gram is discussed. In the Skip-gram model, the weight of words farther away from the current word is smaller, while the weight of words closer to the current word is higher. The reason is that the analysis unit is small, so the amount of training data can be increased, the training performance is also better, and it also has the advantage of processing low-frequency words. The framework used in the present invention utilizes the Word2Vec model for protein sequences proposed by Asgari and Mofrad et al., called ProtVec, and the Word2Vec model for drug simplified molecular linear input specification (SMILES) sequences proposed by Öztürk1 and Ozrimli et al., called SMILEsVec. The frameworks of SMILEsVec and ProtVec both belong to the Skip-gram model.

In step 102, a training/prediction result generation action is performed; the processing result is introduced into a prediction affinity model module 3 for training, where these representations are combined with protein-ligand docking interaction features and placed into the deep model of the prediction affinity model module 3 for prediction; and a prediction result generation action is performed to generate a prediction result through the training of the prediction affinity model module 3, and then proceed to step 103.

Here, in step 102, the processing result is introduced into a prediction affinity model module 3 for training, so as to generate a prediction result through the training of the prediction affinity model module 3, that is, the sequence data and the structural data are introduced into the prediction affinity model module 3 which can be a convolutional neural model/deep neural model for training to generate a prediction result, that is, the drug vector and the protein vector generated by the above data processing are processed to generate a prediction result, and are integrated with the drug protein structure data D_drug_protein.

Among them, a deep learning model for predicting affinity model module 3 will be constructed to quickly and accurately predict the interaction (affinity) between drugs and targets (proteins). The size of the interaction between drugs and targets (proteins) is obtained by using 39 features obtained from Autodock and the representation of protein and drug sequence data learned from the Word2vec model. The established model is compared with other related studies and other machine learning models, and the cross-validation method is used to select the best performing model parameters, and the test set is used to obtain data results.

In the case of deep neural networks, problems such as overfitting and long training time may occur during training. Hinton proposed Dropout to solve this problem. During the training process, some neurons in the hidden layer are temporarily discarded with a certain probability, so that the neurons are multiplied by 0 to stop computing. In this way, the network will become thinner, which can solve the problem of time-consuming network. In addition, when the network is being trained, the input of each layer will be affected by the parameters of the previous layer, and the input parameter distribution will continue to change during the training process. In this way, each layer must adapt to the new distribution, making the network training more complicated. This change in data distribution during training is called internal covariate shift. This problem covers not only adjusting the hyperparameters of the model, but also initializing parameters and adjusting the learning rate. To solve this problem, S. Ioffe proposed the batch normalization (BN) method in 2015. The BN method takes regularization as part of the network model architecture and uses it in each mini-batch. When the data is forward-propagated, each layer is normalized, the mean and standard deviation of the distribution are modulated to 0 and 1, and finally the data is expanded and translated.

In the context of the present invention, in any model architecture used, the BN layer is placed before the activation function layer.

In the case of the convolutional neural network (CNN), in the CNN architecture used in the present invention, the dimension of the convolutional base layer used is one-dimensional. The one-dimensional convolutional neural network is suitable for time series analysis. For example, in the protein and drug data sets used, the protein amino acid sequence and the drug simplified molecular linear input norm (SMILEs) are both one-dimensional, and features are learned and extracted using a one-dimensional convolutional layer. The convolutional layer is a set of parallel feature maps. For example, in terms of images, the convolution learns image features using small blocks of input data to retain the relationship between pixels. It is composed of sliding different convolution kernels on the input image and performing certain operations. This is a mathematical operation that requires two inputs, such as an image matrix and a filter or kernel, so that each element in a feature map is derived by repeatedly performing a convolution.

In terms of multimodal learning, the present invention uses multimodal learning that is biased towards multimodal representation learning. Single-modal representation learning represents data as mathematical vectors that can be processed by the model or further abstracts higher-level feature vectors. In many studies and applications, deep learning models are very successful in supervised learning of single-type data representations, while multimodal representation learning uses multimodality to learn better feature representations in the data and combines the features of each modality so that the model can better learn the data. In the present invention, in addition to the feature data under protein-drug interaction, there is also data of protein sequence and drug string. Multimodal learning has been successfully applied to integrating video and audio in deep learning models or successfully applied to speech synthesis in past studies. However, few studies have combined interaction features and protein target sequence data through deep learning to predict binding affinity. Here, the present invention combines a DNN (for interaction features) and a CNN (for protein sequence and drug string). This method is to merge the final output layers of the two models and further utilize multiple hidden layers in series when confirming the final prediction. For multimodal learning, integrating different data types can benefit from different aspects.

In step 103, a sorting operation is performed; the prediction results are sorted, and the mean square error D_MSE generated by the prediction affinity model module is transmitted to the sorting model module, and then the sorting model module is used to sort and obtain the consistency index D_CI. Here, the prediction results are sorted, that is, the result of integrating the prediction results with the drug protein structure data D_drug_protein is introduced into the sub-network of the deep network model, and the generated multi-modal architecture model is then introduced into the prediction affinity model module 3 formed by the sub-network to process and generate the mean square error D_MSE, and then sorted by the sorting model module 4 and obtain the consistency index D_CI. In this another embodiment, the data processing is to process the drug data and protein data according to their data types, and the processing results generated are at least one sequence data and at least one structure data, and the predicted affinity model module 3 is a deep network model. However, the above is only used for example, and the present invention is not limited to the implementation method of the above another embodiment.

Here, in the evaluation model method of the present invention, two methods are used to detect and compare the prediction capabilities of different models. These two methods are the consistency index D_CI (index of Concordance, C-index) and the mean squared error D_MSE (mean squared error, MSE).

As for the consistency index D_CI, since the network model predicts the interaction affinity under the protein-ligand interaction, and the affinity is a continuous value, the present invention uses the consistency index (CI) as an evaluation indicator of prediction accuracy. Mathematically speaking, the CI on a set of paired data is the probability that the prediction of two randomly selected drug-ligand pairs with different label values is performed in the correct order.

In terms of mean square error D_MSE, the mean square error is used to observe the expected value of the square of the difference between the true label value and the predicted value. The smaller the value, the more accurate the model prediction.

In the present invention, in order to evaluate the model's ability to predict and rank drug-target interactions, random forest (RF) and gradient boost (GB) are used to compare with the model established in the present invention.

FIG. 3 is a schematic diagram showing the system architecture and operation status of an embodiment of the deep learning model system of the present invention.

The data processing service module 2, the predicted affinity model module 3, and the ranking model module 4 are at least one of hardware, firmware, and software, and operate in conjunction with a processor (not shown) of an electronic device where the deep learning model system 1 is located. Here, the electronic device may be, for example, a personal computer, a server, an AI server, an Android phone, an iPhone, a tablet computer, an iPad, etc.

Predictive affinity model module 3, data processing service module 2 imports the processing result into the predictive affinity model module 3 for training, so as to generate a prediction result through the training of the predictive affinity model module 3, that is, the sequence data and the structural data are imported into the predictive affinity model module 3 which can be a convolutional neural model/deep neural model for training to generate a prediction result, that is, the drug vector and protein vector generated by the above data processing are processed to generate a prediction result, and integrated with the drug protein structure data D_drug_protein.

The prediction results are sorted, and the results of integrating the prediction results with the drug protein structure data D_drug_protein are introduced into the sub-network of the deep network model. The generated multi-modal architecture model is then introduced into the prediction affinity model module 3 formed by the sub-network for processing to generate the mean square error D_MSE.

The sorting model module 4 receives the mean square error D_MSE from the predicted affinity model module 3, and sorts and obtains the consistency index D_CI through the sorting model module 4; the data processing is to process the drug data and protein data according to their data types, and the processing results generated are at least one sequence data and at least one structure data, and the predicted affinity model module 3 is a deep network model.

The data processing service module 2 processes at least one drug data (drug string data (SMILES)) and protein data (protein sequence data (Protein Sequence, for example, one-dimensional protein sequence features)) to generate at least one processing result corresponding to the at least one drug data and protein data; the data processing service module 2 processes at least one drug data and protein data to generate at least one processing result corresponding to the at least one drug data and protein data, wherein the drug data and protein data include the drug's SMILES (Simplified molecular input line entry specification, SMILES) D_drug 12 and protein sequence D_protein 11 and the drug protein structure data D_drug_protein 13, the data processing can be generated by executing the data processing service module of the natural language processing Word2vec model, and the processing result is at least one sequence data, and is integrated with the drug protein structure data D_drug_protein 13.

Here, Word2vec is used to learn the at least one protein sequence data feature (Protein Sequence, for example, one-dimensional protein sequence feature) and the drug string data SMILES (ligand SMILES string (at least one drug string data (ligand simplified molecular linear input specification (SMILES))). Word2vec is applied to one of the deep learning natural language processing (NLP) methods, and uses natural language methods to explore and deepen the understanding of life language. The low-dimensional feature vectors of these protein sequences and drug strings are learned and obtained according to the grammar encoded in the protein sequence data feature amino acid sequence and the drug string data SMILES.

Each data processed by the data processing service module 2 includes three types: protein-ligand docking interaction features, protein sequences, and ligand simplified molecular linear input specifications SMILES (i.e., drug data (drug string data)), which are obtained from different platforms; protein-ligand docking interaction features are extracted by simulation by Autodock Vina software, for example, 39 protein-ligand docking interaction feature names. Protein sequences are ProteinIDs in the data that are compared and obtained using the UniProt database. UniProt is a comprehensive database of protein sequences and functions, many of which come from genome projects and contain a large amount of information on the biological functions of proteins from research literature and many studies. The SMILES uses the CID in the data to compare with the PubChem database and obtain the sequence. The PubChem database provides biological activity data of small organic molecules related to biological activities. It is maintained by the National Center for Biotechnology Information (NCBI), a subsidiary of the National Institutes of Health (NIH). PubChem is part of the NIH "Molecular Libraries Roadmap Initiative". Information can be accessed for free at www. You can also use FTP to download millions of chemical databases for free. The PubChem database can query various organic chemical molecules, including drugs and chemical raw materials.

The drug data includes the ligand (drug) simplified molecular linear input specification SMILES D_drug 12, and the protein sequence data (protein data) includes the protein sequence D_protein 11, and the protein ligand docking interaction characteristics include drug protein structure data D_drug_protein 13. The data processing of the data processing service module 2 can be generated by executing the natural language processing Word2vec model to generate the processing result, and the processing result is at least one sequence data, and is integrated with the drug protein structure data D_drug_protein13 of the protein ligand docking interaction characteristics.

Data processing service module 2 uses the technical features of Word2vec and uses Word2vec (a related model used to generate word vectors (word2vec)) to learn the one-dimensional protein sequence and ligand SMILES string. Applied to deep learning natural language processing (NLP) methods; in the present invention, in addition to the amino acid characteristics and chemical formula information in the primary structure (sequence) of the protein and the drug string, the three-dimensional structure of the protein and the drug also has extremely important information that allows the model to learn more comprehensive drug-target docking features. Therefore, the protein (target) 3D structure and the automatically generated binding pattern of the drug (ligand) in the docking simulation software can be used to calculate and obtain the structure-based interaction characteristics between the protein and the drug. These features are combined with other data types, such as one-dimensional protein sequence data and drug SMILES strings, and a method for combining these data types to predict and establish the binding affinity of protein-drug interactions for deep learning models is proposed. To this end, the present invention uses the Word2vec model for generating word vectors to learn one-dimensional protein sequences and ligand SMILES strings. The Word2vec model for generating word vectors is applied to one of the deep learning natural language processing (NLP) methods. The existing natural language methods are used to explore and deepen the understanding of life language. The Word2vec model, one of the deep learning natural language processing (NLP) methods, is applied to learn one-dimensional protein sequences and drug SMILES strings. The grammar encoded in the protein amino acid sequence and the SMILES string is followed to learn and obtain these sequences and strings mapped to feature vectors, and the converted vectors are combined with the interaction features and input into the prediction affinity model module 3 for prediction.

The present invention is different from the above-mentioned prior art and literature. Considering that traditional docking simulation software such as Autodock Vina, GOLD, ZDOCK, RDOCK, and HADDOCK can predict and sort docking points, docking simulation software can also predict binding affinity. However, docking simulation software simplifies many calculation formulas due to the consideration of calculation time cost, resulting in many important information for affinity not being considered, such as factors such as water solvent or hydrogen bond strength, resulting in the docking simulation software not having a very good performance in predicting binding affinity. Although it can be combined with molecular dynamics simulation (MD simulation) or Quantum-Mechanics (QM) to increase the accuracy, this must take a lot of time and cost.

In the present invention, the protein 3D structure and the automatically generated binding pattern of the ligand in the docking simulation software are used to calculate and obtain the structure-based interaction characteristics between the protein and the ligand. These characteristics are combined with other data types, such as one-dimensional protein sequence data and ligand SMILES strings, and a method for combining these data types to predict the binding affinity of protein-ligand interaction with a deep learning model is proposed. For this purpose, Word2vec (a related model used to generate word vectors (word2vec)) is used to learn the one-dimensional protein sequence and ligand SMILES string. Word2vec is applied to one of the deep learning natural language processing (NLP) methods, and the existing natural language methods are used to explore and deepen the understanding of the language of life. According to the protein amino acid sequence and SMILES The grammar encoded in the strings is learned to obtain low-dimensional feature vectors of these sequences and strings, and these representations are combined with interaction features and then put into the deep model for prediction.

Affinity prediction model module 3, wherein the data processing service module 2 utilizes the technical features of Word2vec, and uses Word2vec (a related model (word2vec) for generating word vectors) to learn protein sequence data, for example, a one-dimensional protein sequence, and a string of SMILES of a ligand (drug) simplified molecular linear input specification, Word2vec is applied to a deep learning natural language processing (NLP) method; in the present invention, in addition to the amino acid characteristics and chemical formula information in the primary structure (sequence) of the protein and the drug string, the three-dimensional structure of the protein and the drug also has extremely important information that allows the affinity prediction model module 3 to learn more comprehensive drug-target docking features (protein-ligand docking interaction features).

In addition, the prediction affinity model module 3 and the data processing service module 2 import the processing result into the prediction affinity model module 3 for training, so as to generate a prediction result through the training of the prediction affinity model module 3. That is to say, the sequence data and the structural data are imported into the prediction affinity model module 3 which can be a convolutional neural model/deep neural model for training to generate a prediction result, that is, the drug vector and the protein vector generated by the above data processing are processed to generate a prediction result, and are integrated with the drug protein structure data D_drug_protein 13.

Here, using the data processing service module 2 and/or the affinity prediction model module 3, the protein (target) 3D structure and the automatically generated binding pattern of the drug (ligand) in the docking simulation software can be used to calculate and obtain the structure-based interaction characteristics (protein-ligand docking interaction characteristics) between the protein and the drug, and these characteristics (protein-ligand docking interaction characteristics) are combined with other data types, such as one-dimensional protein sequence data and drug SMILES strings, and a method for combining these data types to predict and establish a deep learning model for protein-drug interaction binding affinity is proposed. For this purpose, the present invention uses the relevant model Word2 used to generate word vectors vec is used to learn one-dimensional protein sequences and ligand SMILES strings. The related model used to generate word vectors is applied to one of the deep learning natural language processing (NLP) methods. Through existing natural language methods, we explore and deepen the understanding of life language. The Word2vec model, one of the deep learning natural language processing (NLP) methods, is applied to learn one-dimensional protein sequences and drug SMILES strings. The grammar encoded in the protein amino acid sequence and SMILES string is learned and mapped to feature vectors. The converted vectors are combined with the interaction features and input into the prediction affinity model module 3 for prediction.

In addition, in the affinity prediction model module 3, it is also proposed to use Multi-Modal to learn these data representations and interaction characteristics and predict binding affinity, and compare them with common machine learning models such as Random Forest and Gradient Boost. The results show that the performance of combining different types of data, whether it is Mean Square Error (MSE) or Concordance Index (C-index), is 10~15% higher than the results of using only one type of data. It is hoped that by combining more types of data, more accurate and better ranked drug-target interactions can be predicted, and a platform can be established to perform multiple analyses on many protein families and polypharmacology to explore unknown characteristics.

The target used in the technical features of the present invention is to explore the inhibitory effect of enzyme activity, and the enzyme is serine hydrolase. Serine hydrolase is one of the largest and most diverse enzyme species known in organisms. In mammals, it accounts for about 1% of all proteins, and in humans, there are about 240 members. Serine hydrolases play many key roles in biological processes, such as blood homeostasis, neural signaling and glucose homeostasis in pathological and physiological processes. They also have important functions in viruses and bacteria, such as bacterial and viral infections, life cycles, toxicity and drug resistance. Among the more than 200 physiological compounds and functions of human serine hydrolases, many do not have good characteristics and most lack selective active inhibitors. We try to conduct high-throughput, multiplex analysis of serine hydrolases, explore their characteristics, and create an analysis platform. We use 96 serine hydrolase members and 52 drugs to conduct experimental observations and record the inhibition concentration and the characteristics after interaction with each other, so as to improve the selectivity of inhibitors for more human serine hydrolases.

The prediction affinity model module 3 imports at least one of the processing results into the prediction affinity model module 3 for training, so as to generate a prediction result through the training of the prediction affinity model of the prediction affinity model module 3, that is, the sequence data and the structure data are imported into the prediction affinity model module 3 which can be a convolutional neural model/deep neural model for training to generate a prediction result, that is, the drug vector and the protein vector generated by the above data processing are processed to generate the prediction result, and the prediction result generated is integrated with the drug protein structure data D_drug_protein 13 using the data processing service module 2 and/or the prediction affinity model module 3.

The sorting model module 4 is used to sort the prediction results; the prediction results are integrated with the drug protein structure data D_drug_protein 13 and the results are introduced into the sub-network of the deep network model. Here, for example, the predicted affinity model module 3 is a deep network model, and the generated multi-modal architecture model is then introduced into the predicted affinity model module 3 formed by the sub-network to be processed to generate the mean square error D_MSE, and the predicted affinity model module transmits D_MSE to the sorting model module 4; then, the sorting is performed through the sorting model module 4 to obtain the consistency index D_CI.

Here, in one embodiment, the data processing is to process the drug data and protein data according to their data types, and the processing results generated are at least one sequence data and at least one structure data, and the predicted affinity model module 3 is a deep network model.

Data processing service module 2 utilizes the technical features of Word2vec and uses Word2vec (a related model (word2vec) used to generate word vectors) to learn one-dimensional protein sequences and ligand SMILES strings. Word2vec is applied to deep learning natural language processing (NLP) methods. In the present invention, in addition to the amino acid characteristics and chemical formula information in the primary structure (sequence) of the protein and the drug string, the three-dimensional structure of the protein and the drug also has extremely important information that allows the model to learn more comprehensive drug-target docking features. Therefore, the protein (target) 3D structure and the automatically generated binding mode of the drug (ligand) in the docking simulation software can be used to calculate and obtain the structure-based interaction characteristics between the protein and the drug.

In the present invention, the protein 3D structure and the automatically generated binding pattern of the ligand in the docking simulation software are used to calculate and obtain the structure-based interaction characteristics between the protein and the ligand. These characteristics are combined with other data types, such as one-dimensional protein sequence data and ligand SMILES strings, and a method for combining these data types to predict the binding affinity of protein-ligand interaction with a deep learning model is proposed. For this purpose, Word2vec (a related model used to generate word vectors (word2vec)) is used to learn the one-dimensional protein sequence and ligand SMILES string. Word2vec is applied to one of the deep learning natural language processing (NLP) methods, and the existing natural language methods are used to explore and deepen the understanding of the language of life. According to the protein amino acid sequence and SMILES The grammar encoded in the strings is learned to obtain low-dimensional feature vectors of these sequences and strings.

FIG. 4 is a flow chart illustrating a process step of performing a deep learning model method using an embodiment of the deep learning model system of the present invention as shown in FIG. 3 .

In step 201, data processing is performed; the data processing service module 2 performs data processing on at least one drug data (drug string data (ligand simplified molecular linear input specification (SMILES)) and protein data (protein sequence data (Protein Sequence, for example, one-dimensional protein sequence characteristics)) to generate at least one processing result corresponding to the at least one drug data and protein data, respectively. Here, Word2vec is used to learn the at least one protein sequence data (Protein Sequence, for example, one-dimensional protein sequence characteristics) and the ligand SMILES string (at least one drug string data (ligand simplified molecular linear input specification (SMILES)), Word2vec One of the deep learning natural language processing (NLP) methods is applied, and a natural language method is used to explore and deepen the understanding of the language of life. The low-dimensional feature vectors of these sequences and strings are learned and obtained according to the grammar encoded in the protein amino acid sequence and the SMILES string, and then proceed to step 202.

Here, in step 201, the data processing service module 2 performs data processing on at least one drug data and protein data respectively, and generates at least one processing result corresponding to the at least one drug data and protein data respectively. The drug data and protein data here include the simplified molecular input line entry specification (SMILES) D_drug 12 of the drug, the protein sequence D_protein 11 and the drug protein structure data D_drug_protein 13. The data processing can be performed by the data processing service module 2 executing the natural language processing (word2vec) model to generate the processing result, and the processing result is at least one sequence data, and is integrated with the drug protein structure data D_drug_protein 13.

As for Word2Vec, the present invention uses the skip-gram model, so only Skip-gram is discussed. In the Skip-gram model, the weight of words farther away from the current word is smaller, while the weight of words closer to the current word is higher. The reason is that the analysis unit is small, so the amount of training data can be increased, the training performance is also better, and it also has the advantage of processing low-frequency words. The framework used in the present invention utilizes the Word2Vec model for protein sequences proposed by Asgari and Mofrad et al., called ProtVec, and the Word2Vec model for drug simplified molecular linear input specification (SMILES) sequences proposed by Öztürk1 and Ozrimli et al., called SMILEsVec. The frameworks of SMILEsVec and ProtVec both belong to the Skip-gram model.

In step 202, a training/prediction result generation action is performed; the processing result is introduced into a prediction affinity model module 3 for training, where these representations are combined with protein-ligand docking interaction features and placed in the deep model of the prediction affinity model module 3 for prediction; and a prediction result generation action is performed to generate a prediction result through the training of the prediction affinity model module 3, and then proceed to step 203.

Here, in step 202, the processing result is introduced into a prediction affinity model module 3 for training, so as to generate a prediction result through the training of the prediction affinity model module 3, that is, the sequence data and the structural data are introduced into the prediction affinity model module 3 which can be a convolutional neural model/deep neural model for training to generate a prediction result, that is, the drug vector and the protein vector generated by the above data processing are processed to generate a prediction result, and are integrated with the drug protein structure data D_drug_protein 13.

Among them, a deep learning model for predicting affinity model module 3 will be constructed to quickly and accurately predict the interaction (affinity) between drugs and targets (proteins). The size of the interaction between drugs and targets (proteins) is obtained by using 39 features obtained from Autodock and the representation of protein and drug sequence data learned from the Word2vec model. The established model is compared with other related studies and other machine learning models, and the cross-validation method is used to select the best performing model parameters, and the test set is used to obtain data results.

In the case of deep neural networks, problems such as overfitting and long training time may occur during training. Hinton proposed Dropout to solve this problem. During the training process, some neurons in the hidden layer are temporarily discarded with a certain probability, so that the neurons are multiplied by 0 to stop computing. In this way, the network will become thinner, which can solve the problem of time-consuming network. In addition, when the network is being trained, the input of each layer will be affected by the parameters of the previous layer, and the input parameter distribution will continue to change during the training process. In this way, each layer must adapt to the new distribution, making the network training more complicated. This change in data distribution during training is called internal covariate shift. This problem covers not only adjusting the hyperparameters of the model, but also initializing parameters and adjusting the learning rate. To solve this problem, S. Ioffe proposed the batch normalization (BN) method in 2015. The BN method takes regularization as part of the network model architecture and uses it in each mini-batch. When the data is forward-propagated, each layer is normalized, the mean and standard deviation of the distribution are modulated to 0 and 1, and finally the data is expanded and translated.

In the context of the present invention, in any model architecture used, the BN layer is placed before the activation function layer.

In the case of the convolutional neural network (CNN), in the CNN architecture used in the present invention, the dimension of the convolutional base layer used is one-dimensional. The one-dimensional convolutional neural network is suitable for time series analysis. For example, in the protein and drug data sets used, the protein amino acid sequence and the drug simplified molecular linear input norm (SMILEs) are both one-dimensional, and features are learned and extracted using a one-dimensional convolutional layer. The convolutional layer is a set of parallel feature maps. For example, in terms of images, the convolution learns image features using small blocks of input data to retain the relationship between pixels. It is composed of sliding different convolution kernels on the input image and performing certain operations. This is a mathematical operation that requires two inputs, such as an image matrix and a filter or kernel, so that each element in a feature map is derived by repeatedly performing a convolution.

In terms of multimodal learning, the present invention uses multimodal learning that is biased towards multimodal representation learning. Single-modal representation learning represents data as mathematical vectors that can be processed by the model or further abstracts higher-level feature vectors. In many studies and applications, deep learning models are very successful in supervised learning of single-type data representations, while multimodal representation learning uses multimodality to learn better feature representations in the data and combines the features of each modality so that the model can better learn the data. In the present invention, in addition to the feature data under protein-drug interaction, there is also data of protein sequence and drug string. Multimodal learning has been successfully applied to integrating video and audio in deep learning models or successfully applied to speech synthesis in past studies. However, few studies have combined interaction features and protein target sequence data through deep learning to predict binding affinity. Here, the present invention combines a DNN (for interaction features) and a CNN (for protein sequence and drug string). This method is to merge the final output layers of the two models and further utilize multiple hidden layers in series when confirming the final prediction. For multimodal learning, integrating different data types can benefit from different aspects.

In step 203, a sorting action is performed; the prediction results are sorted, that is, the prediction results and the drug protein structure data D_drug_protein are integrated into the sub-network of the deep network model, and the generated multi-modal architecture model is then introduced into the prediction affinity model module 3 formed by the sub-network to be processed to generate the mean square error D_MSE, and then sorted by the sorting model module 4 to obtain the consistency index D_CI. In this other embodiment, the data processing is to process the drug data and protein data according to their data types, and the processing results generated are at least one sequence data and at least one structure data, and the prediction affinity model module 3 is a deep network model, but the above is only used for example, and the present invention is not limited to the implementation method of the above another embodiment.

Here, in the evaluation model method of the present invention, two methods are used to detect and compare the prediction capabilities of different models. These two methods are the consistency index D_CI (index of Concordance, C-index) and the mean squared error D_MSE (mean squared error, MSE).

As for the consistency index D_CI, since the network model predicts the interaction affinity under the protein-ligand interaction, and the affinity is a continuous value, the present invention uses the consistency index (CI) as an evaluation indicator of prediction accuracy. Mathematically speaking, the CI on a set of paired data is the probability that the prediction of two randomly selected drug-ligand pairs with different label values is performed in the correct order.

In terms of mean square error D_MSE, the mean square error is used to observe the expected value of the square of the difference between the true label value and the predicted value. The smaller the value, the more accurate the model prediction.

In the present invention, in order to evaluate the model's ability to predict and rank drug-target interactions, random forest (RF) and gradient boost (GB) are used to compare with the model established in the present invention.

The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of the present invention; any other equivalent changes or modifications that are accomplished without departing from the spirit disclosed by the present invention should be included in the following patent scope.

It is obvious to those skilled in the art that the present invention can be implemented in other specific forms without departing from the spirit of the present invention. Therefore, the above description should not be interpreted as limiting in all aspects, but should be interpreted as illustrative.

The scope of the present invention should be determined by reasonable interpretation of the appended claims, and all changes within the scope of equivalents of the present invention are included in the scope of the present invention.

1: Deep learning model system 2: Data processing service module 3: Prediction affinity model module 4: Ranking model module 11: D_protein 12: D_drug 13: D_drug_protein 101 102 103: Steps 201 202 203: Steps

Figure 1 is a system schematic diagram for showing the system architecture and operation of the deep learning model system of the present invention; Figure 2 is a flow chart for showing the process steps of using the deep learning model system of the present invention as shown in Figure 1 to perform a deep learning model method; Figure 3 is a schematic diagram for showing the system architecture and operation of an embodiment of the deep learning model system of the present invention; Figure 4 is a flow chart for showing the process steps of using an embodiment of the deep learning model system of the present invention as shown in Figure 3 to perform a deep learning model method.

101 102 103: Steps

Claims (8)

A deep learning model method is applied in the environment of drug-target interaction prediction and ranking, and includes the following procedures: Performing data processing actions; performing data processing on at least one drug data and protein data respectively, and generating at least one processing result corresponding to the at least one drug data and protein data respectively, wherein the at least one drug data is a SMILES drug string data, and the protein data is a protein sequence data, and the data processing is performed by using a protein (target) 3D structure in a docking simulation software The structure and the automatically generated binding mode of the drug (ligand) are used to calculate and obtain a structure-based interaction feature between the protein and the drug, and the interaction feature is combined with the protein sequence data and the ligand simplified molecular linear input specification (SMILES) drug string data to generate the at least one processing result by executing the natural language processing Word2vec model, wherein the interaction feature is a protein-ligand docking interaction feature; Perform training/generate prediction results; import the processing results into a prediction affinity model module for training; and generate prediction results, learn and obtain the protein sequence data and the SMILES drug string data according to the grammar encoded in the protein sequence data and the SMILES drug string data to map to feature vectors, and combine the converted feature vectors with the interaction features to generate a prediction result through the training of the prediction affinity model module; and sort the prediction results. The deep learning model method as described in item 1 of the patent application scope, wherein the predicted affinity model module is a deep network model. A deep learning model method as described in claim 1, wherein the natural language processing Word2vec model is applied to a deep learning natural language processing (NLP) method. The deep learning model method as described in item 1 of the patent application scope, wherein the processing result is introduced into the prediction affinity model module for training, so as to generate the prediction result through the training of the prediction affinity model module. A deep learning model system is applied in the environment of drug-target interaction prediction and ranking, comprising: a data processing service module; a prediction affinity model module; and a ranking model module; wherein the data processing service module performs data processing on at least one drug data and protein data respectively, and generates at least one processing result corresponding to the at least one drug data and protein data respectively, wherein the at least one drug data is a SMILES drug string data, and the protein data is a protein sequence data, and the data processing of the data processing service module is performed by using a protein in a docking simulation software ( The target) 3D structure and the automatically generated binding pattern of the drug (ligand) are used to calculate and obtain a structure-based interaction feature between the protein and the drug, and the interaction feature is combined with the protein sequence data and the ligand simplified molecular linear input specification (SMILES) drug string data to generate at least one processing result by executing the natural language processing Word2vec model, wherein the interaction feature is a protein-ligand docking interaction feature; The processing result is introduced into the prediction affinity model module for training; and the prediction result is generated by learning and mapping the protein sequence data and the SMILES drug string data to feature vectors according to the grammar encoded in the protein sequence data and the SMILES drug string data, and the transformed feature vector is combined with the interaction feature to generate a prediction result through the training of the prediction affinity model module; and The prediction result is sorted, and the mean square error D_MSE generated by the processing of the prediction affinity model module is transmitted to the sorting model module, and then sorted by the sorting model module to obtain the consistency index D_CI. A deep learning model system as described in item 5 of the patent application scope, wherein the predicted affinity model module is a deep network model. A deep learning model system as described in Item 5 of the patent application, wherein the natural language processing Word2vec model is applied to a deep learning natural language processing (NLP) method. A deep learning model system as described in Item 5 of the patent application scope, wherein the processing result is introduced into the prediction affinity model module for training to generate the prediction result through the training of the prediction affinity model module.

Images