Imprinted polymer property prediction method based on graph attention transfer learning
Through graph attention transfer learning and weighted multi-task learning, the problem of low efficiency in the development of imprinted polymers is solved, and efficient molecular feature extraction and improved prediction accuracy are achieved. It is suitable for fields such as sensor development, immunological analysis and environmental monitoring.
Patent Information
- Application Number
- CN202510797977.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-13
- Publication Date
- 2025-09-16
AI Technical Summary
Traditional imprinted polymer synthesis methods are costly and complex to operate. Traditional machine learning relies on manually designed features and requires a large amount of high-quality data. However, the field of molecularly imprinted polymers lacks a large database, resulting in inefficient development of imprinted polymers.
A graph attention transfer learning method is used to construct a high-quality source dataset by screening molecules with similar shapes and sizes. A graph attention network is used to extract molecular features, and knowledge transfer and weighted multi-task learning are performed in the target domain to optimize model performance.
It improves the development efficiency of imprinted polymers, enhances the performance on scarce data sets, and improves the prediction accuracy and fitting ability of the model.
Smart Images
Figure CN120656590A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence-designed imprinted polymers, and in particular to a method for predicting properties of imprinted polymers based on graph attention transfer learning. Background Art
[0002] Molecularly imprinted polymers (MIPs) are synthesized in the presence of specific template molecules. They are highly selective, stable, and reusable, and are widely used in fields such as sensor development, immunological analysis, and environmental monitoring. However, traditional trial-and-error synthesis methods have many drawbacks. They are not only costly and cumbersome, but also pose a complex, high-dimensional optimization challenge for parameters such as the ratio of functional monomers, crosslinker selection, and polymerization conditions, making them difficult to effectively explore using traditional experimental methods.
[0003] Machine learning, and graph neural networks in particular, show significant potential for improving the efficiency of imprinted polymer development. However, traditional machine learning typically relies on manually designed features and requires expert experience to extract features such as molecular topology, functional groups, bond lengths, and bond angles, placing high demands on technical expertise. Graph neural networks greatly simplify this process, directly inputting molecular graphs and automatically extracting local features within them through operations like convolution and pooling, ultimately generating molecular-level features, demonstrating a higher level of intelligence. The attendant problem is that graph neural networks require large amounts of high-quality data to learn effective features and make more accurate judgments. However, there is currently no such large-scale database in the field of molecularly imprinted polymers. Therefore, efficiently utilizing existing data and expanding high-quality data is a major challenge in the intelligent development of imprinted polymers. Summary of the Invention
[0004] Purpose of the invention: The purpose of the present invention is to provide a method for predicting properties of imprinted polymers based on graph attention transfer learning, which is used to improve the development efficiency of imprinted polymers.
[0005] Technical solution: The method for predicting imprinted polymer properties based on graph attention transfer learning described in the present invention comprises the following steps:
[0006] (1) Screen molecules based on molecular shape and size from public datasets to construct high-quality source datasets;
[0007] (2) Perform molecular characterization on the source and target datasets to obtain a molecular feature matrix;
[0008] (3) Build a graph attention network model and train the model using the source dataset until convergence;
[0009] (4) Transfer the knowledge learned by the source domain training model to the target domain, and fine-tune the model on the target dataset. At the same time, use weighted multi-task learning to help the model focus on key tasks.
[0010] Furthermore, the step (1) screens the molecules in the public quantum chemistry dataset whose molecular shapes and sizes are similar to the template molecules in the target dataset to produce the source dataset, and the shape and size similarity in the screening conditions are balanced between the number of molecules and the molecular mass, thereby maximizing the gain of transfer learning.
[0011] Furthermore, the process of screening the source data set in step (1) includes:
[0012] (1.1) Use the RDKit tool library to calculate the molecular volume and maximum atomic distance of the template molecule;
[0013] (1.2) Traverse the molecules in the QM9 dataset one by one and calculate the molecular volume and maximum atomic distance of each molecule;
[0014] (1.3) Screen molecules that meet the conditions of volume difference and maximum interatomic distance difference with the template molecule as the source data set.
[0015] Furthermore, step (2) uses the RDKit library to extract atomic features and bond features from the SMILES string, and generates a molecular feature matrix for input into the graph neural network through one-hot encoding, which is expressed as:
[0016] F molecule =Concat(F atom , F bond )
[0017] Among them, F molecule represents the molecular feature matrix, F atom represents the atomic feature matrix, F bond Represents the key feature matrix.
[0018] Furthermore, step (3) constructs a graph attention network in the source domain to extract molecular features, and the fully connected layer predicts molecular properties based on the generated graph embedding.
[0019] Furthermore, the step (3) includes:
[0020] (3.1) The He method is used to initialize the model parameters. The model weight W starts from a mean of 0 and a standard deviation of Sampling from the normal distribution of , the mathematical expression is:
[0021]
[0022] Among them, nis the number of input units in the current layer, and W is the model parameter;
[0023] (3.2) The model obtains the output value through forward propagation, as shown in the attention network:
[0024]
[0025] in, is the feature representation of node v at the kth layer; is the set of neighbor nodes of node v; is the attention coefficient of the kth layer, obtained through learning; W( k ) is the weight matrix of the kth layer; σ is the activation function;
[0026] (3.3) Define the loss function to calculate the loss. The regression task uses the mean square error MSE:
[0027]
[0028] Among them, n is the number of samples, yi is the true value, is the predicted value;
[0029] (3.4) Minimizing the loss helps the model converge faster. The gradient descent optimization algorithm is combined with MSE:
[0030]
[0031] After multiple rounds of iterations of the above process until convergence, it is migrated to the target task.
[0032] Furthermore, the loss function in step (3) adopts MSE, and the Adam optimizer is used for training optimization. The early stopping mechanism is adopted during the training process, the threshold is 50, and the training is performed for a maximum of 500 rounds. That is, if the R 2 If the value does not improve for 50 consecutive rounds, stop training. After training is completed, use R 2 Indicators such as RMSE are used to evaluate the model's fitting ability and prediction accuracy.
[0033] Furthermore, the step (4) includes:
[0034] (4.1) Loading the pre-trained model parameters into the target task model, using the general quantum chemistry knowledge learned in the source domain as prior knowledge to achieve knowledge transfer;
[0035] (4.2) Reconstruct the model network structure to adapt to multi-task learning in the target domain. The single-task architecture of the source model is adjusted to a multi-task architecture. Multi-task learning optimizes multiple related tasks simultaneously. The calculation process is as follows:
[0036]
[0037] Where N is the number of tasks, w i and are the weight and loss of the i-th task, is the total loss;
[0038] (4.3) Fine-tuning the model, including freezing some layers, fine-tuning the entire model, and unfreezing layer by layer. The learning error of knowledge transfer is composed of the following three parts:
[0039]
[0040] Among them, ∈ T (h) is the expected error on the target domain; ∈ S (h) is the training error on the source domain; is the degree of distribution difference between the two domains; λ is the error difference of the optimal model between the two domains.
[0041] Furthermore, the loss function in step (4) adopts MSE, and the Adam optimizer is used for training optimization. The early stopping mechanism is adopted during the training process, the threshold is 30, and the training is carried out for a maximum of 300 rounds. That is, if the R 2 If the value does not improve for 30 consecutive rounds, the training is stopped. After the training is completed, the model's fitting ability and prediction accuracy are evaluated using indicators such as R2 and RMSE.
[0042] Furthermore, the weighted multi-task learning in step (4) assigns different task weights according to the importance of the tasks, and assigns more weights to tasks that more significantly affect the binding affinity of the imprinted polymer.
[0043] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:
[0044] (1) This paper proposes to solve the data scarcity problem of imprinted polymers through transfer learning, and proposes to use molecular shape and size similarity to control the quality of source data during the transfer process to ensure efficient knowledge transfer;
[0045] (2) This paper proposes to use attention networks to process graph data. The attention mechanism can focus on key information, thereby enhancing the performance of the model on scarce datasets.
[0046] (3) This paper proposes weighted multi-task learning, which helps the model focus on core tasks by assigning different loss weights to different tasks, thereby improving the performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 It is a framework flow chart of the present invention;
[0048] Figure 2 Schematic diagram of the dataset screening method. DETAILED DESCRIPTION
[0049] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0050] like Figure 1 The figure shows a framework for imprinted polymer property prediction based on graph attention transfer learning. First, high-quality molecular data is screened from public datasets to create a source dataset, which is then converted into graph data for attention network training. The resulting graph embedding data is fed into a fully connected layer for property prediction. Once the model is fully trained in the source data domain, the model weight parameters are used to initialize the model parameters in the target domain to achieve knowledge transfer. The fully connected layer is then rebuilt in the target domain to perform similar model training as described above. This specifically includes the following four aspects:
[0051] (1) Select data filter
[0052] The template molecule in the target dataset (SMILES string COc1ccccc1C(Cc2ccccc2)N3CCCCC3) is used as the benchmark molecule, which theoretically has the best binding performance with MIPs. Subsequently, molecules similar in shape and size to the template molecule are screened from the QM9 dataset as candidate molecules for the source dataset, such as Figure 2 Specific screening criteria include volume difference and maximum interatomic distance difference. The screening criteria are set as follows: volume difference within ±30% / ±20% / ±10%, and maximum interatomic distance difference within ±35% / ±25% / ±15%. The screening criteria gradually increase the similarity between the source and target data, from loose to strict. The specific algorithm flow is shown in Table 1. This screening method ensures the similarity between the source and target data, allowing the common features and molecular structure patterns learned by the model from the source data to be effectively applied to the target task.
[0053] Table 1: Source data screening algorithm process
[0054]
[0055]
[0056] (2) Molecular characterization
[0057] Molecules need to be processed into a suitable graph data structure before they can be processed by graph neural networks. Molecular descriptors are numerical representations of molecular structures that quantify the properties of atoms and bonds within a molecule. Using the RDKit library, atomic features and bond features can be extracted from SMILES strings and generated through one-hot encoding to generate a molecular feature matrix for input into graph neural networks. This process can be expressed as: F molecule =Concat(Fatom , F bond ). Among them, F molecule is the molecular feature matrix, F atom is the atomic feature matrix, F bond is the bond feature matrix. One-hot encoding converts discrete eigenvalues into binary vectors, avoiding the potential interference of feature value differences on model learning and ensuring that the model treats each feature category equally. In some embodiments, the molecular features involved (atomic features and bond features) and their respective value ranges are shown in Table 2.
[0058] Table 2: Molecular feature set and its value range
[0059]
[0060] (3) Source domain pre-training
[0061] The process typically involves:
[0062] (3.1) The He method is used to initialize the model parameters (weights and biases). The model weight W starts with a mean of 0 and a standard deviation of Sampling from the normal distribution of , the mathematical expression is:
[0063]
[0064] Among them, n in is the number of input units in the current layer, and W is the model parameter;
[0065] (3.2) The model obtains the output value through forward propagation, as shown in the attention network:
[0066]
[0067] in, is the feature representation of node v at the kth layer; is the set of neighbor nodes of node v; is the attention coefficient of the kth layer, obtained through learning; W (k) is the weight matrix of the kth layer; σ is the activation function;
[0068] (3.3) Define the loss function to calculate the loss. The regression task uses the mean square error MSE:
[0069]
[0070] Where n is the number of samples, y i is the true value, is the predicted value;
[0071] (3.4) Minimizing the loss helps the model converge faster. The gradient descent optimization algorithm is combined with MSE:
[0072]
[0073]
[0074] After multiple rounds of iterations of the above process until convergence, it is migrated to the target task.
[0075] In some embodiments, the specific settings are: the number of hidden layers of the model is set to 3, the number of output tasks is set to 1, the molecular feature dimension is 75, and the initial learning rate is 0.001. The loss function uses MSE, and the Adam optimizer is used for training optimization. The early stopping mechanism (threshold is 50) is used during the training process, and a maximum of 500 rounds of training are performed. That is, if the R 2 If the value does not improve for 50 consecutive rounds, stop training. 2 Indicators such as RMSE are used to evaluate the model's fitting ability and prediction accuracy.
[0076] (4) Target domain fine-tuning
[0077] Generally includes the following steps:
[0078] (4.1) Load the pre-trained model parameters into the target task model, and use the general quantum chemistry knowledge learned in the source domain as prior knowledge to achieve knowledge transfer.
[0079] (4.2) Restructure the model network structure to adapt to multi-task learning in the target domain, such as adjusting the single-task architecture of the source model to a multi-task architecture. Multi-task learning optimizes multiple related tasks simultaneously, and it is necessary to comprehensively consider the loss of each task. The calculation process is as follows:
[0080]
[0081] Where N is the number of tasks, w i and are the weight and loss of the i-th task, is the total loss; classic multi-task learning assigns equal weight to each task, that is
[0082] In some embodiments, the substrate binding amount S b Directly quantifying the degree of binding between MIPs and substrates is the core indicator for measuring binding affinity; the free concentration C fAlthough it can indirectly reflect the binding equilibrium state, its role in affinity assessment is relatively weak. The classic equal multi-task learning method cannot meet the above requirements. Based on the early exploratory experiments, unequal weights are assigned to the two tasks, and different task weights are assigned according to the importance of the tasks: The total loss calculation process is as follows:
[0083]
[0084] (4.3) Fine-tuning the model, including freezing some layers, fine-tuning the entire model, and unfreezing layer by layer. The learning error of knowledge transfer is composed of the following three parts:
[0085]
[0086] Among them, ∈ T (h) is the expected error on the target domain; ∈ S (h) is the training error on the source domain; is the degree of distribution difference between the two domains; λ is the error difference of the optimal model between the two domains.
[0087] In some embodiments, the specific settings are: the number of hidden layers of the model is set to 2, the number of output tasks is set to 2, the molecular feature dimension is 75, and the initial learning rate is 0.001. The loss function uses MSE, and the Adam optimizer is used for training optimization. The early stopping mechanism (threshold is 30) is used during training, and a maximum of 300 rounds of training are performed. That is, if the R 2 If the value does not improve for 30 consecutive rounds, stop training. 2 Indicators such as RMSE are used to evaluate the model's fitting ability and prediction accuracy.
[0088] This paper proposes to address the data scarcity issue of imprinted polymers through transfer learning. During the transfer process, it uses molecular shape and size similarity to control the quality of the source data, thereby ensuring efficient knowledge transfer. Secondly, this paper proposes using attention networks to process graph data. The attention mechanism can focus on key information, thereby enhancing the model's performance on scarce datasets. Furthermore, this paper proposes weighted multi-task learning, which helps the model focus on core tasks by assigning different loss weights to different tasks, thereby improving model performance.
Claims
1. A method for predicting imprinted polymer properties based on graph attention transfer learning, characterized in that: The steps include: (1) Screen molecules based on molecular shape and size from public datasets to construct high-quality source datasets; (2) Perform molecular characterization on the source and target datasets to obtain a molecular feature matrix; (3) Build a graph attention network model and train the model using the source dataset until convergence; (4) Transfer the knowledge learned by the source domain training model to the target domain, and fine-tune the model on the target dataset. At the same time, use weighted multi-task learning to help the model focus on key tasks.
2. The imprinted polymer property prediction method based on graph attention transfer learning according to claim 1 is characterized in that: The step (1) screens the public quantum chemistry dataset for molecules whose molecular shapes and sizes are similar to the template molecules in the target dataset to produce a source dataset, and the shape and size similarity in the screening conditions are balanced between the number of molecules and the molecular mass, thereby maximizing the gain of transfer learning.
3. The imprinted polymer property prediction method based on graph attention transfer learning according to claim 1, characterized in that: The screening process of the source data set in step (1) includes: (1.1) Use the RDKit tool library to calculate the molecular volume and maximum atomic distance of the template molecule; (1.2) Traverse the molecules in the QM9 dataset one by one and calculate the molecular volume and maximum atomic distance of each molecule; (1.3) Screen molecules that meet the conditions of volume difference and maximum interatomic distance difference with the template molecule as the source data set.
4. The imprinted polymer property prediction method based on graph attention transfer learning according to claim 1, characterized in that: The step (2) uses the RDKit library to extract atomic features and bond features from the SMILES string, and generates a molecular feature matrix for input into the graph neural network through one-hot encoding, which is expressed as: F molecule =Concat(F atom ,F bond ) Among them, F molecule represents the molecular feature matrix, F atom represents the atomic feature matrix, F bond Represents the key feature matrix.
5. The imprinted polymer property prediction method based on graph attention transfer learning according to claim 1, characterized in that: The step (3) constructs a graph attention network in the source domain to extract molecular features, and the fully connected layer predicts molecular properties based on the generated graph embedding.
6. The imprinted polymer property prediction method based on graph attention transfer learning according to claim 1, characterized in that: The step (3) comprises: (3.1) The He method is used to initialize the model parameters. The model weight W starts from a mean of 0 and a standard deviation of Sampling from the normal distribution of , the mathematical expression is: Among them, n is the number of input units in the current layer, and W is the model parameter; (3.2) The model obtains the output value through forward propagation, as shown in the attention network: in, is a node v Feature representation of the kth layer; is the set of neighbor nodes of node v; is the attention coefficient of the kth layer, obtained through learning; W( k ) is the weight matrix of the kth layer; σ is the activation function; (3.3) Define the loss function to calculate the loss. The regression task uses the mean square error MSE: Among them, n is the number of samples, yi is the true value, is the predicted value; (3.4) Minimizing the loss helps the model converge faster. The gradient descent optimization algorithm is combined with MSE: After multiple rounds of iterations of the above process until convergence, it is migrated to the target task.
7. The imprinted polymer property prediction method based on graph attention transfer learning according to claim 6, characterized in that: The loss function in step (3) adopts MSE, and the Adam optimizer is used for training optimization. The early stopping mechanism is adopted during the training process, the threshold is 50, and the training is carried out for a maximum of 500 rounds. That is, if the R 2 If the value does not improve for 50 consecutive rounds, stop training. After training is completed, use R 2 Indicators such as RMSE are used to evaluate the model's fitting ability and prediction accuracy.
8. The imprinted polymer property prediction method based on graph attention transfer learning according to claim 1, characterized in that: The step (4) comprises: (4.1) Loading the pre-trained model parameters into the target task model, using the general quantum chemistry knowledge learned in the source domain as prior knowledge to achieve knowledge transfer; (4.2) Reconstruct the model network structure to adapt to multi-task learning in the target domain. The single-task architecture of the source model is adjusted to a multi-task architecture. Multi-task learning optimizes multiple related tasks simultaneously. The calculation process is as follows: Where N is the number of tasks, w i and are the weight and loss of the i-th task, is the total loss; (4.3) Fine-tuning the model, including freezing some layers, fine-tuning the entire model, and unfreezing layer by layer. The learning error of knowledge transfer is composed of the following three parts: Among them, ∈ T (h) is the expected error on the target domain; ∈ S (h) is the training error on the source domain; is the degree of distribution difference between the two domains; λ is the error difference of the optimal model between the two domains.
9. The imprinted polymer property prediction method based on graph attention transfer learning according to claim 8, characterized in that: The loss function in step (4) adopts MSE, and the Adam optimizer is used for training optimization. The early stopping mechanism is adopted during the training process, the threshold is 30, and the training is carried out for a maximum of 300 rounds. That is, if the R 2 If the value does not improve for 30 consecutive rounds, the training is stopped. After the training is completed, the model's fitting ability and prediction accuracy are evaluated using indicators such as R2 and RMSE.
10. The imprinted polymer property prediction method based on graph attention transfer learning according to claim 1, characterized in that: The weighted multi-task learning in step (4) assigns different task weights according to the importance of the tasks, and assigns more weights to tasks that have a more significant impact on the binding affinity of the imprinted polymer.