A knowledge-guided method and system for predicting metal-ion-organic interactions

Abstract

This invention discloses a knowledge-guided method and system for predicting metal-ion-organic interactions, belonging to the fields of environmental separation and computational chemistry. The method constructs a training set through DFT calculations, extracts organic molecule descriptors, and quantifies their prior relationship with binding free energy. It employs a graph attention network to extract molecular features and fuses metal ion features using a cross-attention mechanism. A loss function incorporating prior knowledge is designed, and high-precision prediction is achieved through two-stage training. This invention solves the problem of traditional machine learning's dependence on large amounts of labeled data, maintaining excellent performance even in sparse data scenarios. It is suitable for high-throughput screening of separation agents such as extractants and adsorbents, significantly reducing R&D costs.

Description

Technical Field

[0001] This invention belongs to the field of environmental separation and computational chemistry, specifically relating to a knowledge-guided method and system for predicting metal-ion-organic interactions. Background Technology

[0002] The separation and recovery of metal ions is a key measure to mitigate heavy metal ion pollution and minimize the waste of metal resources, and is of great significance to environmental protection and resource recycling. The formation of stable complexes between metal ions and functional groups (e.g., hydroxyl, amino, carboxyl groups) of organic matter (such as extractants, adsorbents, and membranes) is an important chemical reaction mechanism for metal ion separation and recovery. As a key thermodynamic parameter in this process, the binding free energy of metal ions with organic matter is... This can reveal the spontaneity of complexation reactions and the stability of metal complexes, thereby quantifying the interaction strength between metal ions and organic matter. Based on this, This technology can help researchers understand separation reaction mechanisms at the microscopic molecular level and serve as a key indicator for screening separation agents in environmental separation processes such as extraction and adsorption. Therefore, the ability to efficiently and accurately assess the binding free energy between metal ions and organic matter has significant practical implications.

[0003] Density functional theory (DFT) offers acceptable computation time and high computational accuracy compared to traditional experimental methods and Hartree-Fock first-principles calculations, making it the most widely used method currently. Calculation method. Nevertheless, since the computational complexity of DFT is... ( For structural optimization of larger systems (such as metal complexes), frequency analysis and single-point energy DFT calculations still incur high computational costs. This is especially true considering the structural diversity of organic compounds and the complexity of solvation environments, which necessitate high-throughput computations. For applications such as molecular screening, the computational cost of DFT is extremely high.

[0004] To address the high cost of acquiring target data in the field of chemistry, researchers often choose to establish surrogate models to rapidly acquire key chemical properties / functions data through a data-driven approach. Many machine learning models, especially deep learning models, are widely used in molecular structure-activity / structure-property studies in chemistry due to their excellent ability to solve high-dimensional nonlinear problems. Therefore, establishing data-driven machine learning models to predict the binding free energy of organic compounds with metals based on their fundamental structure is an efficient and accurate way to acquire such data. Effective methods for data analysis. Summary of the Invention

[0005] For existing technologies, machine learning methods are used for prediction. To address the problem of excessively high computational costs caused by the need to collect or construct large amounts of training data, this invention proposes a knowledge-guided method and system for predicting metal-ion-organic interactions. This invention implements a machine learning prediction framework that integrates chemical knowledge guidance and data-driven approaches, making it suitable for implementation in sparse data scenarios. The framework provides high-precision predictions. Therefore, it can be migrated to new systems at a low cost, demonstrating strong applicability.

[0006] The technical solution of the present invention is as follows:

[0007] A knowledge-guided method for predicting metal-ion-organic interactions includes the following steps:

[0008] A training set was constructed by calculating the binding free energy between metal ions and organic matter using density functional theory.

[0009] Molecular descriptors of organic compounds are extracted and their prior relationship with binding free energy is quantified using a random forest model;

[0010] Graph attention network was used to extract features of organic molecules, and multilayer perceptron was used to extract features of metal ions.

[0011] The characteristics of organic matter and metal ions are fused through a cross-attention mechanism;

[0012] Design a loss function that incorporates prior knowledge to guide model training.

[0013] In the above technical solution, the density functional theory calculation uses PBE0-D3 functionals, and the main group elements use 6-31G. The basis set is Stuttgart pseudopotential basis set for transition metals, and def2-TZVP basis set is used for single-point energy calculation.

[0014] In the above technical solution, the molecular descriptor includes electronegativity, molecular weight, surface polarity, highest occupied molecular orbital energy level and lowest unoccupied molecular orbital energy level, and descriptors that are significantly correlated with binding free energy are screened through correlation analysis.

[0015] In the above technical solution, the cross-attention mechanism calculates the attention weights of metal ion features and each atomic feature, and obtains the fused feature by weighted summation of each atomic feature based on the weights.

[0016] In the above technical solution, the loss function includes:

[0017] Losses can be predicted using free energy;

[0018] Molecular descriptor prediction loss;

[0019] Prior knowledge residual loss;

[0020] The total loss is the weighted sum of the losses mentioned above.

[0021] In the above technical solution, the training process is divided into two stages:

[0022] The first phase of simultaneous optimization combines free energy prediction loss, molecular descriptor prediction loss, and prior knowledge residual loss.

[0023] The second stage only optimizes the combination of free energy prediction loss and prior knowledge residual loss.

[0024] The above technical solution also includes a pre-training step:

[0025] Predicting fundamental molecular properties using a graph attention network pre-trained on a large-scale organic dataset;

[0026] Freeze the parameters of the pre-trained layer for subsequent free energy prediction.

[0027] A metal-ion-organic interaction prediction system, comprising:

[0028] The data preprocessing module is used for density functional theory calculations and training set construction;

[0029] The prior knowledge extraction module quantifies the relationship between molecular descriptors and binding free energy based on the random forest model;

[0030] The feature fusion module uses a cross-attention mechanism to integrate the features of organic matter and metal ions, and characterizes the interaction between the two.

[0031] The model training module enables two-stage knowledge-guided training.

[0032] In the above technical solution, the cross-attention weights of the feature fusion module can be used to explain the interaction between metal ions and key atoms of organic matter.

[0033] A computer-readable storage medium storing a program corresponding to the above method for predicting the binding free energy of metal ions to organic compounds.

[0034] Beneficial effects:

[0035] 1. Improved data efficiency: Guided by prior knowledge, the need for training data is reduced, making it suitable for scenarios with scarce data.

[0036] 2. High prediction accuracy: By using graph neural networks to characterize and extract molecular microstructures and using cross-attention mechanisms to capture atomic-level interactions, the MAE is reduced compared to traditional methods.

[0037] 3. High interpretability: Visual attention weights reveal the contribution of key functional groups (such as carboxyl and amino groups) to the binding energy.

[0038] 4. Low transfer cost: Pre-training and knowledge-guided strategies enable the model to quickly adapt to new metal ion systems. Attached Figure Description

[0039] Figure 1 A flowchart of the overall method for predicting metal-ion-organic interactions;

[0040] Figure 2 This is a schematic diagram of molecular structure optimization and energy calculation using DFT calculations.

[0041] Figure 3 This is a schematic diagram illustrating prior knowledge extraction and random forest modeling.

[0042] Figure 4 A feature fusion architecture diagram of GAT and cross-attention mechanism;

[0043] Figure 5 This is a schematic diagram illustrating the two-stage training strategy and loss function optimization. Detailed Implementation

[0044] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. However, the following embodiments are only for explaining the present invention, and the scope of protection of the present invention should include all the contents of the claims. Moreover, through the description of the following embodiments, those skilled in the art can fully implement all the contents of the claims of the present invention.

[0045] Example:

[0046] The knowledge-guided method for predicting metal-ion-organic interactions proposed in this invention mainly includes the following steps, such as: Figure 1 As shown:

[0047] Step 1 ( Figure 1 101 in the middle): High precision DFT calculations were performed on the data to construct the training set. DFT was then used to optimize the structure and calculate single-point energies of candidate metal ions, organic molecules, and metal-organic complexes, referencing... Figure 2 201 and Figure 2 202. Hybrid basis sets are used in structural optimization, i.e., 6-31G is used for the main family elements. The basis set used was the Stuttgart pseudopotential basis set for transition metals; the more accurate def2-TZVP basis set was used for single-point energy calculations; and the PBE0-D3 functional, commonly used in transition metal systems, was selected. Furthermore, the solution environment was simulated using the PCM implicit solvent model, and the Gibbs free energy correction was obtained using the RRHO model. The binding free energy between the metal ions and organic compounds was calculated using the molecular / ionic Gibbs free energy obtained from DFT calculations, combined with the metal-organic complexation reaction equation. .

[0048] Step 2 ( Figure 1 (102) Extraction and quantification of prior chemical knowledge to guide model training. The interaction strength between metal ions and organic ligands is closely related to the intrinsic descriptors of the organic ligands, which is prior chemical knowledge that can guide model training. However, there is currently no explicit mathematical equation that can quantify this relationship at the microscopic level. Therefore, this invention uses statistical methods to extract and quantify prior chemical knowledge from the model training process. Extract this prior chemical knowledge from the dataset, referencing Figure 2 203 in the table. First, this invention used Multiwfn and RDKit to calculate 27 common molecular descriptors, as shown in Table 1. Then, through Pearson correlation analysis, it identified descriptors related to... Several significantly related descriptors Furthermore, a random forest model was established. Metal ion characteristics (See Table 4 in step 3) and Quantitative relationship between .because Revealing organic molecule descriptors and The quantitative relationship between them is considered as prior knowledge in this invention and incorporated as a constraint condition. Predictive models to improve model performance.

[0049] Table 1

[0050]

[0051] Step 3 ( Figure 1 (103) Characterization of organic molecules and metal ions. To capture the intrinsic characteristics of organic molecules and metal ions, this invention employs multimodal feature engineering, referring to... Figure 3301. Due to the complex and diverse structures of organic molecules, and considering the need for high-throughput prediction to avoid time-consuming calculations such as density functional theory (DFT), organic molecules are represented as graph structures: nodes correspond to atoms, and edges correspond to chemical bonds (see Tables 2 and 3). The characteristics of atoms and chemical bonds can be efficiently calculated using Python's DeepChem library. For metal ions with simple structures and limited chemical space, geometric and electronic characteristics obtained from the literature are used. The ionic radius, molecular weight, electronegativity, etc., are characterized as shown in Table 4 and embedded into a one-dimensional vector. This invention employs one-hot encoding for classification features and normalizes numerical features to eliminate dimensional differences.

[0052] Table 2

[0053]

[0054] Table 3

[0055]

[0056] Table 4

[0057]

[0058] Step 4 ( Figure 1 (104) Feature Extraction and Fusion of Organic Molecules and Metal Ions: The feature extraction module extracts more discriminative features from the initial representations of organic molecules and metal ions, referencing... Figure 3 302. For the graph representation of organic molecules, a Graph Attention Network (GAT) is used for feature extraction. GAT learns adaptive attention weights to adjust the influence of neighboring atoms on the features of each atom, effectively representing the inter-atomic relationships within the molecule. Considering that metal ions are represented in one-dimensional vector form, this invention uses a Multilayer Perceptron (MLP) network to obtain effective features. For prediction... To capture their interactions, multimodal features from molecules and metal ions need to be fused. Cross-modal feature fusion is achieved through a graph readout mechanism, which aggregates atomic features into molecular-level features. It's important to note that traditional readout operations only aggregate features from the graph itself and cannot characterize the interactions between organic molecules and metal ions. For example, for different metal ions, the contribution of each atom in the organic molecule to the prediction result may have subtle differences. Therefore, this invention proposes a readout method based on a cross-attention mechanism, which can integrate metal ion features into the molecular representation, referencing... Figure 3 305 in the context. Specifically, for a given... An organic molecule with 10 atoms (excluding hydrogen atoms) can be represented by the following formula based on cross-attention readout. Where Indicates the characteristics of metal ions. Indicates the first The characteristics of an organic atom These are the parameters that can be optimized through model training. Attention weights. Representing the molecular characteristics of each atom The degree of contribution, and pass The weighted sum is obtained. Note that... It is a one-dimensional vector, which can be used with... The splicing results in fusion features .

[0059] (1)

[0060] (2)

[0061] (3)

[0062] In addition, prior knowledge is introduced into model training. Organic molecule descriptor It must also be included as model output. Because Regardless of metal ion characteristics, this invention employs average pooling to extract molecular-level features. For subsequent use The prediction, reference Figure 3 302 in the middle.

[0063] Step 5 ( Figure 1 105): Predicted output of organic matter descriptor and binding free energy. In the target value prediction module, this invention uses two multilayer perceptron (MLP) networks respectively... and As input for prediction and ,refer to Figure 3 303. To prevent overfitting, this invention employs Dropout technology during training to temporarily disable some neurons in the model.

[0064] Step 6 ( Figure 1 (106) Constructing a loss function based on prior knowledge: Integrating prior knowledge into the model is achieved by adding a constraint term to the training loss function. (See reference...) Figure 3 304 in the middle. Total loss. It consists of two parts: data loss and prior knowledge loss. Data loss includes the predicted value (…). and ) and corresponding true value ( and The mean square error (MSE) between the two is denoted as follows: and , minimize and This ensures that the model predictions fit the real data. Prior knowledge residuals. Used for quantification model prediction with prior knowledge The results Minimize the deviation between them. The mean square error of 0 (denoted as ) This ensures the consistency between the model's two predicted outputs and prior knowledge. Therefore, the total loss... As shown in the following formula:

[0065] (4)

[0066] Prior knowledge residual loss Through coefficients The weighting coefficient, used as a hyperparameter, controls the contribution of prior knowledge to model training. Generally, constraints are set. Because of prior knowledge It contains inherent biases from the random forest model, while This can mitigate its negative impact on model training.

[0067] Step 7 ( Figure 1 107): Model training and evaluation. The model training process includes pre-training (see reference 107). Figure 4 401) and knowledge-guided training (reference) Figure 4 The training set consists of two stages (402 in the original text). Due to the high computational cost of DFT, the training set can only contain a limited number of organic molecules, which restricts the model's ability to perceive molecular structural features. To address this issue, we constructed a pre-training layer based on the GAT network, implementing weakly supervised learning on a larger-scale dataset. The pre-training is designed as a multi-task learning framework, predicting basic structural properties (including Log P, number of rings, number of heavy atoms, number of hydrogen bond donors, number of hydrogen bond acceptors, topological polar surface area, proportion of sp3 hybridized carbon atoms, molar refractive index, number of rings, etc.) calculated using the RDKit tool. Figure 5 (Ref. 503). Specifically, molecular features are first extracted using the GAT network in the pre-trained layer (see reference 503). Figure 5 (refer to 501), then use a global average pooling layer and a multilayer perceptron network (MLP) to extract full-image features and output predicted values ​​(see reference). Figure 5 (502 in the middle). A GAT-based pre-trained layer was added to... Figure 3In the 302 organic molecule feature extraction network, its parameters are frozen during subsequent training. Pre-training enhances the model's ability to perceive organic molecule structures. During the knowledge-guided training phase, molecular descriptors are calculated for each training cycle. Loss on the dataset and Loss on the dataset The model parameters are updated using the Adam optimizer. Furthermore, to promote model convergence, we propose a two-stage training strategy: [The text abruptly ends here, so the translation stops as well.] The first phase consists of one training cycle, using the loss function. Enable the model to learn synchronously and The true value and prior knowledge; the second stage starts from the first Starting from the current cycle, only losses are used. Perform parameter optimization. This stage aims to accelerate... Convergence of predicted loss. Predicted value ( ) and tag value ( The mean absolute error (MAE) and coefficient of determination between the two are: This is used to evaluate the model's learning ability and generalization performance. Simultaneously, the prediction uncertainty generated by Monte Carlo Dropout is used to assess prediction reliability and further define the model's applicability.

[0068] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A knowledge-guided method for predicting metal-ion-organic interactions, characterized in that, Includes the following steps: A training set was constructed by calculating the binding free energy between metal ions and organic matter using density functional theory. Molecular descriptors of organic compounds are extracted and their prior relationship with binding free energy is quantified using a random forest model; Graph attention network was used to extract features of organic molecules, and multilayer perceptron was used to extract features of metal ions. By fusing the characteristics of organic matter and metal ions through a cross-attention mechanism, the interaction between the two can be characterized. Design a loss function that incorporates prior knowledge to guide model training.

2. The knowledge-guided method for predicting metal-ion-organic interactions according to claim 1, characterized in that, The density functional theory calculations used the PBE0-D3 functional, and the main group elements were 6-31G. The basis set is Stuttgart pseudopotential basis set for transition metals, and def2-TZVP basis set is used for single-point energy calculation.

3. The knowledge-guided method for predicting metal-ion-organic interactions according to claim 1, characterized in that, The molecular descriptors include electronegativity, molecular weight, surface polarity, highest occupied molecular orbital energy level, and lowest unoccupied molecular orbital energy level. Descriptors that are significantly correlated with binding free energy are selected through correlation analysis.

4. The knowledge-guided method for predicting metal-ion-organic interactions according to claim 1, characterized in that, The cross-attention mechanism calculates the attention weights of metal ion features and each atomic feature, and then performs a weighted summation of each atomic feature based on these weights to obtain the fused feature.

5. The knowledge-guided method for predicting metal-ion-organic interactions according to claim 1, characterized in that, The loss function includes: Losses can be predicted using free energy; Molecular descriptor prediction loss; Prior knowledge residual loss; The total loss is the weighted sum of the losses mentioned above.

6. The knowledge-guided method for predicting metal-ion-organic interactions according to claim 1, characterized in that, The training process is divided into two phases: The first phase of simultaneous optimization combines free energy prediction loss, molecular descriptor prediction loss, and prior knowledge residual loss. The second stage only optimizes the combination of free energy prediction loss and prior knowledge residual loss.

7. The knowledge-guided method for predicting metal-ion-organic interactions according to claim 1, characterized in that, It also includes a pre-training step: Predicting fundamental molecular properties using a graph attention network pre-trained on a large-scale organic dataset; Freeze the parameters of the pre-trained layer for subsequent free energy prediction.

8. A system for predicting metal-ion-organic interactions, characterized in that, include: The data preprocessing module is used for density functional theory calculations and training set construction; The prior knowledge extraction module quantifies the relationship between molecular descriptors and binding free energy based on the random forest model; The feature fusion module integrates organic and metal ion features using a cross-attention mechanism. The model training module enables two-stage knowledge-guided training.

9. The metal-ion-organic interaction prediction system according to claim 8, characterized in that, The cross-attention weights of the feature fusion module can be used to explain the interaction between metal ions and key atoms in organic matter.

10. A computer-readable storage medium, characterized in that, The program is stored in any of the methods of claims 1-7, for predicting the binding free energy of metal ions to organic compounds.

Images