Framework model based on neural network learning and application method thereof

Abstract

The invention relates to the technical field of medical science, in particular to a framework model based on neural network learning and an application method of the framework model. According to the model, a graph convolutional network (GCN, Transform architecture and long short-term memory network (LSTM) are fused, a multi-level analysis process is constructed, the spatial structure and sequence dynamics of drug molecules are deeply explored, and the model comprises the following specific steps: model training: performing model training by adopting data sets collected by people such as Cui, Q.Cui2020Improved and the like; the model test selects three independent test sets to comprehensively evaluate the performance of the model; according to the data preprocessing, an RDKit library is used, SMILES character strings are converted into molecular objects, a feature tensor is constructed and used for storing core features of each molecule in detail, data splitting is carried out, and the data splitting adopts versions, such as Ahmad, divided by people based on a Cui2020 data set to carry out training and testing of a machine learning model.

Description

Technical Field

[0001] The present invention relates to the field of medical science and technology, and in particular to a framework model based on neural network learning and an application method thereof. Background Art

[0002] In the field of modern medical science, the research and development and application of organic drugs have immeasurable value for improving public health, especially in the fight against chronic diseases such as cancer, cardiovascular disease, diabetes, etc., organic drugs play a heroic role. For example, certain targeted therapies have become an important part of cancer treatment and have significantly improved the survival rate of patients. However, although the progress of modern science and technology has opened up new avenues for disease treatment, drug research and development is still a challenging road.

[0003] The solubility prediction of drug molecules plays a vital role in ensuring their efficacy and safety. It directly affects the absorption, distribution, metabolism and excretion of drugs in the body (ADME process. Traditionally, solubility prediction methods often find it difficult to fully understand the complexity of molecular structures, resulting in a significant deviation between the predicted results and the actual performance. For example, (DADLC mentioned that when dealing with molecular structures with ring structures, important information about the ring structure in the molecule may be lost or not fully captured.

[0004] To address this challenge, this study introduces an innovative deep learning framework that combines an attention-based transformer network, a long short-term memory network (LSTM), and a graph convolutional network (GCN) to more accurately predict the solubility of drug molecules. Summary of the invention

[0005] The purpose of the present invention is to provide a framework model based on neural network learning and its application method, so as to solve the problem raised in the above background technology that when dealing with molecular structures with ring structures, important information of the ring structure in the molecule may be lost or not fully captured, and the solubility of the drug molecule cannot be predicted more accurately. To achieve the above purpose, the present invention provides the following technical solutions: a framework model based on neural network learning, characterized in that: the model is a YZS-Model model, which integrates a graph convolutional network (GCN, Transformer architecture, and a long short-term memory network (LSTM, constructing a multi-level analysis process, deeply exploring the spatial structure and sequence dynamics of drug molecules, and the specific steps of the model are as follows: Model training, the model training uses the dataset collected by Cui, Q. (Cui2020Improved et al.) for model training; Model testing: three independent test sets were selected to comprehensively evaluate the model performance; Data preprocessing: The data preprocessing converts the SMILES string into a molecule object by using the RDKit library, and constructs a feature tensor (features, which is used to store the core features of each molecule in detail; Data splitting, which uses Ahmad et al.'s version of Cui2020 dataset splitting to train and test machine learning models; Molecular feature extraction, which uses rdkit (CiteDrive2022 tool and member methods provided by rdkit to accurately extract various features of atoms and chemical bonds; Extract the features of organic molecules and abstract them into a graph structure; The input of a GNN is an undirected graph G, and the formula G=(V,E, where V is the set of all nodes in the graph, and E is the set of all edges directly connected to the nodes. H. Message passing, the message passing is to calculate the message from each neighbor node u∈N(v) of node v to v, using the formula muv(t+1)=M(hv, hu, xu, xv, euv), where M is the message function, hv and hu are the hidden states of nodes v and u in the tth round, xv and xu are the initial features of the nodes, euv is the feature of the edge connecting v and u, and N(v represents the set of neighbor nodes of v; Aggregation, where node v aggregates the messages received from all its neighbors to generate a single aggregate message, according to the formula av = A ({muv: u∈Nv)}, where A is an aggregation function, which can be summation, average, maximum, etc. State update, the state update makes the update function update the hidden state of the node according to the formula h(t+1)=U(hv(t),av(t+1),xv, where U is the update function which is a simple linear transformation plus nonlinear activation or a more complex neural network; Through the above steps, GNN can update the hidden state of each node in the graph in each round of iteration. Finally, the hidden state of the node will integrate and reflect the feature information of its neighboring nodes and the structural information of the entire graph. After the iterative process continues for several rounds, the final hidden state h is mapped to the predicted label through a readout layer R. According to the formula Yv=R(hv(T), T is the final number of iterations, ýv is the predicted label of node v; Introducing the Graph Convolutional Network (GCN architecture) to process graph data; The Self-Attention Transformer analysis and prediction framework is used to analyze and predict the solubility of organic drugs; Long short-term memory network (LSTM) was used to further process and analyze the solubility of organic drugs; This method uses random feature zeroing technology to explore the specific impact of atomic features on the prediction bias of chemical molecules. For 66 elements, the one-hot encoding method is used to rank the atomic differences independently of other features, so that the individual contribution of each part to the prediction bias can be more accurately evaluated.

[0006] This method adopts an averaging method for the calculation of randomly set zero difference values, that is, taking the average of the feature and its sub-features as the representative value of the total feature. The purpose is to integrate the information of each sub-feature, so as to more comprehensively show the comprehensive role of each feature and the one-hot encoded atom in the model prediction performance.

[0007] This method classifies the properties of organic molecules into atomic types, molecular complexity and structural properties. In order to verify the importance index given by LIME, the solubility index of the top fifteen of the two organic compounds is calculated to qualitatively judge the solubility, and the solubility of pyridine and benzene is explained and predicted.

[0008] This method introduces further expansion of the model and explores other machine learning strategies including semi-supervised learning and unsupervised learning methods to use unlabeled data for model training, improve the model's generalization ability for unknown data, and introduce a multi-scale attention mechanism to more finely capture the complex interactions within molecules and improve the model's prediction accuracy.

[0009] The application of deep learning technology in this method can assist in preclinical drug screening and optimize the chemical properties of candidate drugs to increase their bioavailability. In the field of personalized medicine, this model integrates the individual's genetic background to predict drug response, thereby guiding the formulation of personalized treatment plans.

[0010] Compared with the prior art, the present invention has the following beneficial effects: In the present invention, the framework model based on neural network learning and its application method provide a new solution path for deep learning technology in the context of current drug development, which can significantly improve Accuracy and efficiency of water solubility prediction.

[0011] In the present invention, the framework model based on neural network learning and its application method, automatic feature learning, and deep learning model can automatically learn and extract useful features from complex data without relying on manually selected physical and chemical parameters, which enables it to process extremely complex molecular structures.

[0012] In the present invention, the framework model based on neural network learning and its application method have the ability to process big data. With the rapid growth of scientific research data, deep learning can effectively process and analyze large-scale data sets, including those obtained through high-throughput experiments, thereby improving the generalization ability and accuracy of the prediction model.

[0013] In the present invention, the framework model based on neural network learning and its application method have the adaptability and flexibility of the model, and the deep learning model can optimize the prediction of specific types of molecules by adjusting the network structure and parameters.

[0014] In the present invention, the framework model based on neural network learning and its application method are quickly iterated and updated. The deep learning model can quickly learn from new experimental data and adapt to new research findings and market needs, thereby accelerating the drug development cycle. This can not only accelerate the drug development process, but also help identify candidate molecules that may have good therapeutic potential in the early stages, greatly reducing research and development costs and time. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the process structure of the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technical personnel in this field without creative work are within the scope of protection of the present invention.

[0017] A framework model based on neural network learning, the model is the YZS-Model model, which integrates graph convolutional networks (GCN, Transformer architecture, and long short-term memory networks (LSTM, constructing a multi-level analysis process, deeply exploring the spatial structure and sequence dynamics of drug molecules. The specific steps of the model are as follows: A. Model training,The model training uses the dataset collected by Cui, Q. (Cui2020Improved et al.); B. Model testing,Three independent test sets were selected for model testing to comprehensively evaluate the model performance; C. Data Preprocessing,Data preprocessing is done by using the RDKit library to convert SMILES strings into,molecule objects and construct a feature tensor (features,used to store the core features of each molecule in detail; D. Data splitting: Data splitting uses the version of Ahmad et al. based on the Cui2020 dataset for training and testing machine learning models; E. Molecular feature extraction, molecular feature extraction uses rdkit (CiteDrive2022 tool and member methods provided by rdkit to accurately extract various features of atoms and chemical bonds; F. After extracting the features of organic molecules, abstract them into a graph structure. G. The input of a GNN is an undirected graph G, using the formula G=(V,E, where V is the set of all nodes in the graph and E is the set of all edges directly connected to the nodes. H. Message passing, message passing is to calculate the message passed from each neighbor node u∈N(v) of node v to v, using the formula muv(t+1)=M(hv, hu, xu, xv, euv), where M is the message function, hv and hu are the hidden states of nodes v and u in the tth round, xv and xu are the initial features of the nodes, euv is the feature of the edge connecting v and u, and N(v represents the set of neighbor nodes of v; I. Aggregation,Aggregation is when a node v aggregates the messages received from all its neighbors to generate a single aggregate message, according to the formula av = A ({muv: u∈Nv)}, A is the aggregation function, which can be sum, average, maximum and other operations; J. State update, the state update makes the update function update the hidden state of the node, according to the formula h(t+1)=U(hv(t),av(t+1),xv, where U is the update function which is a simple linear transformation plus nonlinear activation or a more complex neural network; K. Through the above steps, GNN can update the hidden state of each node in the graph in each round of iteration. Finally, the hidden state of the node will integrate and reflect the feature information of its neighboring nodes and the structural information of the entire graph. After several rounds of iteration, the final hidden state h is mapped to the predicted label through a readout layer R. According to the formula Yv=R(hv(T), T is the final number of iterations, ýv is the predicted label of node v; L. Introduce Graph Convolutional Network (GCN architecture to process graph data; M. used the Self-Attention Transformer analysis and prediction architecture to analyze and predict the solubility of organic drugs; N. used a long short-term memory network (LSTM) to further process and analyze the solubility of organic drugs.

[0018] This method uses random feature zeroing technology to explore the specific impact of atomic features on the prediction bias of chemical molecules. For 66 elements, the one-hot encoding method is used to rank the atomic differences independently of other features, so that the individual contribution of each part to the prediction bias can be more accurately evaluated.

[0019] This method adopts an averaging method for the calculation of randomly set zero difference values, that is, taking the average of the feature and its sub-features as the representative value of the total feature. The purpose is to integrate the information of each sub-feature, so as to more comprehensively show the comprehensive role of each feature and the one-hot encoded atom in the model prediction performance.

[0020] This method classifies the properties of organic molecules into atomic types, molecular complexity and structural properties. In order to verify the importance index given by LIME, the solubility index of the top fifteen of the two organic compounds is calculated to qualitatively judge the solubility, and the solubility of pyridine and benzene is explained and predicted.

[0021] This method introduces further expansion of the model and explores other machine learning strategies including semi-supervised learning and unsupervised learning methods to use unlabeled data for model training, improve the model's generalization ability for unknown data, and introduce a multi-scale attention mechanism to more finely capture the complex interactions within molecules and improve the model's prediction accuracy.

[0022] The application of deep learning technology in this method can assist in preclinical drug screening and optimize the chemical properties of candidate drugs to increase their bioavailability. In the field of personalized medicine, this model integrates the individual's genetic background to predict drug response, thereby guiding the formulation of personalized treatment plans.

[0023] The above shows and describes the basic principles, main features and advantages of the present invention. Technical personnel in this industry should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A framework model based on neural network learning, characterized by: This model is YZS-Model The model combines graph convolutional networks (GCNs, Transformer architectures, and long short-term memory networks (LSTMs) to build a multi-level analysis process that deeply explores the spatial structure and sequence dynamics of drug molecules. The specific steps of the model are as follows: A. Model training: The model training uses the data set collected by Cui, Q. Cui 2020 Improved et al. for model training; B. Model testing: three independent test sets were selected to comprehensively evaluate the model performance; C. Data preprocessing: The data preprocessing converts the SMILES string into a molecule object by using the RDKit library, and constructs a feature tensor (features, which is used to store the core features of each molecule in detail; D. Data splitting, which uses Ahmad et al.'s version based on the Cui2020 dataset to train and test machine learning models; E. Molecular feature extraction, which uses the rdkitCiteDrive2022 tool and the member method provided by rdkit to accurately extract various features of atoms and chemical bonds; F. Extract the features of organic molecules and abstract them into a graph structure; G. The input of a GNN is an undirected graph G, using the formula G=V,E, where V is the set of all nodes in the graph and E is the set of all edges directly connected to the nodes; H. Message passing: for node v, the message passing is to calculate the message passed to v from each neighbor node u∈Nv of node v, using the formula muv(t+1)=M(hv, hu, xu, xv, euv), where M is the message function, hv and hu are the hidden states of nodes v and u in the tth round, xv and xu are the initial features of the nodes, euv is the feature of the edge connecting v and u, and Nv represents the set of neighbor nodes of v; I. Aggregation: The aggregation is that node v aggregates the messages received from all its neighbors to generate a single aggregate message, according to the formula av = A{muv:u∈Nv}, where A is the aggregation function, which can be summation, average, maximum value and other operations; J. State update, the state update makes the update function update the hidden state of the node according to the formula h(t+1)=Uhv(t),av(t+1),xv, where U is the update function which is a simple linear transformation plus nonlinear activation or a more complex neural network; K. Through the above steps, GNN can update the hidden state of each node in the graph in each round of iteration. Finally, the hidden state of the node will integrate and reflect the feature information of its neighboring nodes and the structural information of the entire graph. After several rounds of iteration, the final hidden state h is mapped to the predicted label through a readout layer R. According to the formula Yv=Rhv(T), T is the final number of iterations, and ýv is the predicted label of node v; L. Introduce Graph Convolutional Network (GCN architecture to process graph data; M. used the Self-Attention Transformer analysis and prediction architecture to analyze and predict the solubility of organic drugs; N. used a long short-term memory network (LSTM) to further process and analyze the solubility of organic drugs.

2. A neural network learning application method, characterized in that: This method uses random feature zeroing technology to explore the specific impact of atomic features on the prediction bias of chemical molecules. For 66 elements, the one-hot encoding method is used to rank the atomic differences independently of other features, so that the individual contribution of each part to the prediction bias can be more accurately evaluated.

3. A neural network learning application method, characterized in that: This method adopts an averaging method for the calculation of randomly set zero difference values, that is, taking the average of the feature and its sub-features as the representative value of the total feature. The purpose is to integrate the information of each sub-feature, so as to more comprehensively show the comprehensive role of each feature and the one-hot encoded atom in the model prediction performance.

4. A neural network learning application method, characterized in that: This method classifies the properties of organic molecules into atomic types, molecular complexity and structural properties. In order to verify the importance index given by LIME, the solubility index of the top fifteen of the two organic compounds is calculated to qualitatively judge the solubility, and the solubility of pyridine and benzene is explained and predicted.

5. A neural network learning application method, characterized in that: This method introduces further expansion of the model and explores other machine learning strategies including semi-supervised learning and unsupervised learning methods to use unlabeled data for model training, improve the model's generalization ability for unknown data, and introduce a multi-scale attention mechanism to more finely capture the complex interactions within molecules and improve the model's prediction accuracy.

6. A neural network learning application method, characterized in that: The application of deep learning technology in this method can assist in preclinical drug screening and optimize the chemical properties of candidate drugs to increase their bioavailability. In the field of personalized medicine, this model integrates the individual's genetic background to predict drug response, thereby guiding the formulation of personalized treatment plans.

Images