A Machine Learning-Based Method for Predicting Molecular Solvation Free Energy

Abstract

This invention discloses a machine learning-based method for predicting the solvation free energy of molecules, belonging to the field of computational chemistry, and solving the technical problem of the high computational resource and time consumption of molecular dynamics methods. The method includes using the SMILES formulas of solute and solvent molecules as input data; converting the SMILES formulas into one-dimensional, two-dimensional, and three-dimensional molecular descriptors using the RDKit tool; training a machine learning model based on the correlation between the one-dimensional, two-dimensional, and three-dimensional molecular descriptors and experimental solvation free energy using an open-source experimental dataset, thus obtaining a prediction model; and inputting the molecular descriptors of solute and solvent molecules into the prediction model to obtain the predicted solvation free energy of the solute molecule. This invention replaces existing molecular dynamics simulations with a machine learning model, and simultaneously uses machine learning algorithms and experimental data for model training, improving the accuracy and stability of the prediction results.

Description

Technical Field

[0001] This invention relates to the field of computational chemistry, and more particularly to a method for predicting molecular solvation free energy based on machine learning. Background Technology

[0002] Molecular solvation free energy (MSE) is an important area in chemistry, materials science, and biology, and methods for calculating MSE have been widely used in drug design, materials development, and biomolecular simulation. Currently, the calculation of MSE primarily relies on molecular dynamics methods. This method, based on classical force field theory, simulates the interactions of molecules in a solvent through molecular dynamics to calculate their solvation free energy. However, due to the computational complexity of molecular dynamics methods and the limitations of the models themselves, the calculated MSE results often contain errors compared to experimental values. These errors may stem from factors such as the complexity of molecular topology, the inaccuracy of force field parameters, and insufficient handling of solvent effects.

[0003] To reduce these errors, optimization of relevant parameters is usually required. However, this optimization process typically consumes significant computational resources and time, and even after optimization, the improvement in computational accuracy remains limited, especially in complex molecular systems. This limits the practical application of molecular dynamics methods.

[0004] In recent years, machine learning technology has developed rapidly. Machine learning methods can learn from a large amount of data and capture the complex relationship between molecular features and their physicochemical properties, and obtain more accurate prediction results. There is an urgent need for a machine learning method that combines molecular descriptors to calculate and predict solvation free energy.

[0005] In the process of realizing this invention, the inventors discovered at least the following problems in the prior art:

[0006] Calculating the solvation free energy of molecules using molecular dynamics methods requires a significant amount of computational resources and time, and the results are not accurate enough, while also being costly. Summary of the Invention

[0007] The purpose of this invention is to provide a machine learning-based method for predicting molecular solvation free energy, addressing the problems of existing molecular dynamics methods that require significant computational resources and time to calculate molecular solvation free energy, resulting in inaccurate results and high costs. The preferred technical solutions provided by this invention offer numerous advantages, which are detailed below.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] This invention provides a machine learning-based method for predicting the solvation free energy of molecules, comprising the following steps: S100: using the SMILES formulas of solute and solvent molecules as input data; S200: converting the SMILES formulas into one-dimensional, two-dimensional, and three-dimensional descriptors of molecules using the RDKit tool; S300: training a machine learning model based on an open-source experimental dataset by the correlation between the one-dimensional, two-dimensional, and three-dimensional descriptors of molecules and experimental solvation free energy, thereby obtaining a prediction model; S400: inputting the molecular descriptors of solute and solvent molecules into the prediction model to obtain the predicted solvation free energy of the solute molecule.

[0010] Preferably, in step S200, the one-dimensional descriptor includes molecular weight and number of bonds, the two-dimensional descriptor includes molecular graph features, topological polar surface area (TPSA), and molecular connectivity index, and the three-dimensional descriptor includes molecular surface area and molecular volume.

[0011] Preferably, in step S200, the missing values ​​in the SMILES-style transformation are estimated and filled using the K-nearest neighbor classification algorithm, so that the datasets of the one-dimensional descriptors, two-dimensional descriptors, and three-dimensional descriptors remain complete.

[0012] Preferably, in step S200, the datasets of the one-dimensional descriptor, two-dimensional descriptor, and three-dimensional descriptor are further normalized to create a consistent scale for the entire dataset.

[0013] Preferably, in step S300, the open-source experimental data comes from the FreeSolv database, and the molecules in the FreeSolv database include a training set and a test set. The training set is used to train the machine learning model, and the test set is used to verify the prediction model.

[0014] Preferably, in step S300, the machine learning model is trained using two strategies: a first training strategy and a second training strategy. The first training strategy is trained based on the offset of the experimental value directly predicted by molecular features, while the second training strategy is trained based on the difference between the predicted experimental value and the simulated value.

[0015] Preferably, in step S300, the machine learning model includes support vector machine (SVM), random forest (RF), deep neural network (DNN), multiple linear regression (MLR), and extreme gradient boosting (XGB), and the optimal hyperparameter configuration of each model is identified through Bayesian optimization methods.

[0016] Preferably, in step S300, during the training of the machine learning model, an ensemble method and a dimensionality reduction method are also used for model training; in the ensemble method, the prediction result of the prediction model is obtained by combining the predicted outputs of each machine learning model on average; the dimensionality reduction method uses principal component analysis (PCA) algorithm to perform feature dimensionality reduction in order to improve model training efficiency.

[0017] Preferably, the solvent molecule is a polar solvent, a nonpolar solvent, or an ionic solvent, and the solute molecule is a small organic molecule.

[0018] Preferably, the solvent is water, and the solute is an alcohol, ester, ketone, aromatic compound, or a compound containing nitrogen, sulfur, or oxygen functional groups.

[0019] Implementing one of the above-described technical solutions of the present invention has the following advantages or beneficial effects:

[0020] This invention replaces existing molecular dynamics simulations with machine learning models, significantly reducing computation time and making it suitable for high-throughput screening of large-scale molecular libraries. Furthermore, the use of advanced machine learning algorithms and large-scale experimental data for model training significantly improves the accuracy and stability of prediction results while reducing costs, providing a multi-faceted and efficient solution for the rapid prediction of solvation free energy of molecules. Moreover, based on SMILES formulas and molecular descriptors, it avoids the cumbersome molecular dynamics simulation process of traditional methods, simplifying the calculation steps for solvation free energy. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0022] Figure 1 This is a flowchart of a method for predicting molecular solvation free energy based on machine learning, according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, various exemplary embodiments described below will be referenced to the accompanying drawings, which form part of the exemplary embodiments, illustrating various exemplary embodiments that may be used to implement the present invention. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. It should be understood that they are merely examples of processes, methods, and apparatuses consistent with some aspects of the present invention disclosed as detailed in the appended claims, and other embodiments may be used, or structural and functional modifications may be made to the embodiments listed herein without departing from the scope and spirit of the present invention.

[0024] To illustrate the technical solution described in this invention, specific embodiments are described below, showing only the parts related to the embodiments of this invention.

[0025] Example 1:

[0026] like Figure 1As shown, this invention provides a machine learning-based method for predicting molecular solvation free energy, comprising the following steps: S100: using the SMILES (Simplified Molecular Input Line Entry System) formula of solute and solvent molecules as input. The SMILES formula is a simplified molecular linear input specification, which is a specification that explicitly describes molecular structure using ASCII strings. It is a simplified representation method of molecular structure using strings, realizing the digital expression of molecular information. It can be imported into most molecular editing software and converted into two-dimensional graphics or three-dimensional molecular models. It can concisely represent complex molecular structures, enabling computers to obtain more accurate and useful information, and is easy to process and store in computers, thus facilitating machine learning operations. It avoids the complex simulation process and high computational cost in traditional calculations, and greatly improves prediction efficiency. S200: Based on the RDKit tool, SMILES formulas are converted into one-dimensional, two-dimensional, and three-dimensional molecular descriptors. The comprehensive coverage of one-dimensional descriptors (such as molecular weight, number of hydrogen bond donors, and number of hydrogen bond acceptors), two-dimensional descriptors (such as topological polar surface area (TPSA) and molecular connectivity index), and three-dimensional descriptors (such as molecular surface area and volume) encompasses the topological structure, geometric features, and electronic properties of molecules. Due to the inherent complexity of molecular structures, direct prediction using three-dimensional molecular structures often involves complex spatial information and multidimensional data, which machine learning algorithms struggle to effectively capture. To systematically study the relationship between input features and solvent free energy, it is necessary to generate as many features as possible and study different types of features to identify those highly correlated with molecular solvation free energy. Therefore, converting these molecular structures into one-dimensional, two-dimensional, and three-dimensional numerical features suitable for machine learning facilitates more accurate prediction calculations. The RDKit tool provides rich chemical information processing capabilities for processing and analyzing information such as molecular structure, chemical reactions, and chemical properties, facilitating the generation of one-dimensional, two-dimensional, and three-dimensional molecular descriptors. S300: Based on an open-source experimental dataset, a machine learning model is trained by relating one-dimensional, two-dimensional, and three-dimensional molecular descriptors to experimental solvation free energies, resulting in a prediction model. This model can quickly and accurately predict the solvation free energy of target molecules. S400: Molecular descriptors of solute and solvent molecules are input into the prediction model to obtain the predicted solvation free energy of the solute molecule. This invention replaces existing molecular dynamics simulations with a machine learning model, significantly reducing computation time. Statistical molecular dynamics methods are expected to require 500 hours to calculate the solvation free energy of the aforementioned 1000 compounds, while the method of this invention only requires 10 minutes, significantly saving computation time.Therefore, this invention is suitable for high-throughput screening of large-scale molecular libraries. Furthermore, by employing advanced machine learning algorithms and large-scale experimental data for model training, the accuracy and stability of prediction results are significantly improved, while reducing costs. The method provided in this embodiment offers a multi-faceted and efficient solution for the rapid prediction of solvation free energy of molecules. Simultaneously, based on SMILES formulas and molecular descriptors, it avoids the cumbersome molecular dynamics simulation process of traditional methods, simplifying the calculation steps for solvation free energy.

[0027] As an optional implementation, in step S200, the one-dimensional descriptor includes molecular weight and the number of bonds (e.g., the number of hydrogen bond donors and acceptors); the two-dimensional descriptor includes molecular graph features, topological polar surface area (TPSA), and molecular connectivity index; and the three-dimensional descriptor includes molecular surface area and molecular volume. The topological, geometric, and electronic characteristics of the molecule can be obtained through one-dimensional, two-dimensional, and three-dimensional descriptors, thus facilitating the determination of molecular solvation free energy. Further, the descriptors specifically used in this embodiment include APFP, ECFP6, TOPOL, MolProps, and their combined combinations MolPropsAPFP and MolPropsECFP6. A single descriptor, such as APFP, is only one way to describe molecular characteristics; using only this type of descriptor may overlook certain physicochemical features of the molecule, thus potentially leading to an incomplete understanding of molecular characteristics. Combined descriptors, such as MolPropsAPFP, combine two types of molecular descriptors, thereby providing a more comprehensive understanding of molecular characteristics.

[0028] As an optional implementation, in step S200, the missing values ​​in the SMILES transformation are estimated and filled using the K-Nearest Neighbor (KNN) classification algorithm (each sample can be represented by its K nearest neighbors, suitable for automatic classification of large sample sizes). This involves using the KNN algorithm for feature interpolation, ensuring the integrity of the one-dimensional, two-dimensional, and three-dimensional descriptor datasets, thus facilitating machine learning model training. Simultaneously, since machine learning models cannot handle non-numerical features, and some features are non-numerical during the SMILES transformation, deleting these non-numerical features ensures that only relevant numerical data is retained, aiding in effective data analysis. Step S200 also normalizes the one-dimensional, two-dimensional, and three-dimensional descriptor datasets, creating a consistent scale across the entire dataset, resulting in more accurate training results. These two comprehensive preprocessing methods—K-Nearest Neighbor classification and normalization—produce a highly robust dataset, ideal for analysis and subsequent model training.

[0029] As an optional implementation, in step S300, the open-source experimental data comes from the FreeSolv database, preferably version 0.52. FreeSolv is a widely used benchmark dataset for predicting solvation free energy, providing experimental measurements and theoretical calculations of the solvation free energy for 642 small neutral organic molecules. It is an ideal resource for building and evaluating computational models designed to predict solvation properties. The molecules in the FreeSolv database include training and test sets. The training set is used to train the machine learning model, and the test set is used to validate the predictive model. For each individual model, the optimal set of hyperparameters that minimizes the mean free energy (MUE) during cross-validation is first determined. This step ensures that each model is fine-tuned to achieve its best performance on the validation data, balancing the trade-off between model complexity and accuracy. After determining the optimal hyperparameters, the model is retrained using the full training set. This allows the model to fully utilize all available data, which is crucial for ensuring robust predictive performance. After training, the model is applied to the test set, where it generates predictions based on the newly trained configuration. These predictions represent the final output of each model. Of the 642 molecules in the FreeSolv database, 47 were specifically selected for the SAMPL4 blind test, an influential benchmark exercise designed to evaluate and compare different predictive models in the context of solvation and molecular interactions. These 47 molecules were designated as the test set because their free energy values ​​were excluded from the training process to allow for unbiased model validation. This blind test ensures that model performance can be evaluated based on unseen data, simulating real-world predictive scenarios. The remaining 595 molecules from the 642 were used as the training set for developing and optimizing computational models. These molecules provide a rich dataset, enabling the model to learn patterns and relationships between molecular structures and their solvation energies. To avoid data leakage and ensure generalization, no molecules from the test set were used during the model building phase. By separating the molecules in the FreeSolv database into test and training sets, this invention implements best practices in machine learning and statistical analysis, ensuring that reported performance metrics reflect the model's ability to generalize beyond the training data.

[0030] As an optional implementation, in step S300, the machine learning model is trained using two strategies: a first training strategy and a second training strategy. The first training strategy trains the model based on the offset of experimental values ​​directly predicted from molecular features, while the second training strategy trains the model based on the difference between predicted experimental values ​​and simulated values. The results of these two training strategies can be used to verify the advantages of this method compared to existing methods. The first training strategy involves directly predicting experimental values ​​based on molecular features, thereby generating the molecular solvation free energy. The second training strategy incorporates calculated values ​​as additional molecular features, and the model predicts the offset between calculated and experimental values, allowing for correction of the calculated free energy. Based on the first training strategy, the solvation free energy of molecules can be directly predicted. Due to the relatively direct physicochemical relationship between features and results, the results obtained through this method maintain high interpretability. Furthermore, the trained model exhibits strong generalization ability because it focuses on the direct mapping between molecular features and solvation free energy. This characteristic of the model's generalization ability makes it very suitable for predicting the solvation free energy of new molecules, even if they differ from the training data, as long as the molecular feature space remains similar. In this scheme, the preferred specific expression of the first training strategy is: Where A represents any sample in the training set. The calculated free energy was obtained using molecular dynamics (MD) simulations performed using the General Amber Force Field (GAFF). These are the experimental values ​​obtained by the machine learning model in this method. This represents the predicted offset value corresponding to the first training strategy. The second training strategy involves correcting the simulated values ​​by predicting the difference between the experimental and simulated values. This method preserves the correct trend set by the simulation calculations and avoids wasting effort in areas where the calculated values ​​are already highly accurate. This strategy relies heavily on the consistency between the data sources of the training and test sets, especially the computational methods used to generate molecular simulation data. The expression for the second training strategy is: For each training set defined by its descriptor in the first training strategy, a machine learning model is applied using 5x cross-validation, which produces a total of N = 5 trained models, where each individual model predicts its own offset value. The arithmetic mean of these predicted offsets is given, and the standard deviation of the mean can be used as a measure of accuracy.

[0031] As an optional implementation, in step S300, the machine learning models include Support Vector Machine (SVM), Random Forest (RF), Deep Neural Network (DNN), Multiple Linear Regression (MLR), and Extreme Gradient Boosting (XGB). Table 1 defines the hyperparameter space for each model. For small datasets, traditional machine learning algorithms have advantages over complex deep learning methods. They can be trained relatively quickly on standard hardware without requiring large amounts of computational effort, thus significantly reducing computational costs. The aforementioned cluster of model algorithms have different advantages in handling different aspects of prediction tasks. Combining these models facilitates the integration of the advantages of various models, resulting in a more accurate prediction model. For example, Deep Neural Network (DNN) and Extreme Gradient Boosting (XGB) show good consistency for continuous variables, while Random Forest (RF) and Support Vector Machine (SVM) show good consistency for discrete variables. By considering both the physical meaning of features and the intrinsic capabilities of the model, optimal prediction performance can be achieved. The optimal hyperparameter configuration for each model is identified using Bayesian optimization methods. The Bayesian optimization method utilizes the SciKit-Optimize (SKOPT) library, version 0.5.2. It leverages an improved acquisition function, allowing for a more strategic and efficient exploration of the hyperparameter space compared to traditional methods such as random or grid search. By focusing the search on promising regions in the space, Bayesian optimization helps identify high-performance configurations with fewer iterations. This method sets the optimization step size to a maximum of 40 steps, as previous experience suggests that model convergence typically occurs before this limit. For each iteration in the optimization process, the MUE (mean across cross-validation folds) on the validation set is computed and returned to the SKOPT routine. This metric serves as the cost function to guide the selection of the next set of hyperparameters. With each subsequent call, SKOPT further refines its search, aiming to minimize the cost function and achieve optimal performance. This iterative process continues until convergence is observed or the step size limit is reached.

[0032] Table 1. Hyperparameter space definition for each model in the machine learning model.

[0033]

[0034] As an optional implementation, in step S300, during the training of the machine learning model, ensemble and dimensionality reduction methods are also employed for model training. In the ensemble method, the prediction results of the predictive models are obtained by combining the outputs of the average predictions of each machine learning model. The dimensionality reduction method uses Principal Component Analysis (PCA) to perform feature dimensionality reduction, thereby improving model training efficiency. In the ensemble strategy, the predictions of all models are combined by averaging their outputs. This approach leverages the unique strengths of each model, as different models may excel at capturing different patterns or relationships in the data. By integrating their predictions, the ensemble method mitigates the potential weaknesses or biases inherent in any single model. The result is typically an improvement in overall performance, as the ensemble benefits from the diversity of model predictions, thus generalizing better to new data. Finally, the ensemble predictions are compared with the actual values ​​on the test set to calculate the final cross-validation mean (MUE), a metric that provides a comprehensive measure of the ensemble accuracy, indicating the degree of consistency between the combined predictions and the true results. The ensemble learning framework not only enhances the predictive power of the system but also helps reduce the risk of overfitting and ensures stronger performance compared to any single model. Specifically, after determining the best prediction for each individual model, further enhancement is achieved through the ensemble method. By averaging the predictions of all models, a unified set of predictions is generated and then compared with the test set to evaluate the final MUE. The overall predictions are generally closer to the experimental results, effectively reducing the number of outliers. This ensemble approach achieves a higher level of reliability and accuracy in the final results. In model training, a large feature set can introduce noise during training, potentially reducing model accuracy and leading to overfitting. Furthermore, many features may be highly correlated with each other; highly correlated features often express redundant information, either influencing the target label in the same way or having similar physical and chemical properties. This feature redundancy increases computational complexity and may lead to sparsity in the feature matrix, ultimately reducing training efficiency and wasting computational resources. Therefore, feature dimensionality reduction is crucial to the training results. To address the issues of redundant and highly correlated features, Principal Component Analysis (PCA) is applied for dimensionality reduction in this application. The basic principle of PCA is to analyze the correlation between features by calculating the covariance matrix of the data, and then extract the direction of maximum variance (i.e., the principal component direction) through eigenvalue decomposition. These principal components are linear combinations of the original features, ordered by their variance. Components with higher variance capture most of the information in the data, while those with lower variance represent less significant changes and are generally negligible. In implementation, the original feature data is first standardized to ensure that each feature has the same scale, making their contributions to the model equivalent. After standardization, the PCA function from scikit-learn is applied to reduce the dimensionality of the data.During the dimensionality reduction process, a threshold was set for the proportion of variance to be retained. 95% and 75% were chosen as the retention thresholds. This reduces the number of features while retaining as much important information as possible from the original data. The reduced features are then used as input to train the machine learning model. The changes in feature dimension after dimensionality reduction and the corresponding data are shown in Table 2.

[0035] Table 2. Changes in feature dimension after dimensionality reduction and corresponding data for each machine learning model.

[0036]

[0037] As an optional implementation method, the solvent molecules can be polar solvents, non-polar solvents, or ionic solvents, thus making the method more adaptable and able to achieve good prediction results for different types of solvents, which can be widely used in drug development, materials design, and green chemistry.

[0038] As an optional implementation, water is preferably used as the solvent and acetone as the solute in this method. Using the method of the present invention, the predicted value of the solvation free energy of acetone in water is -6.0 kcal / mol, which is only 0.1 kcal / mol less than the experimental value of -6.1 kcal / mol, indicating that the model of the present invention has high prediction accuracy.

[0039] The embodiment is merely a specific example and does not indicate that this is the only way to implement the present invention.

[0040] The above description is merely a preferred embodiment of the present invention. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for predicting molecular solvation free energy based on machine learning, characterized in that, Includes the following steps: S100: Use the SMILES formulas of solute and solvent molecules as input data; S200: Based on the RDKit tool, the SMILES formula is converted into one-dimensional, two-dimensional, and three-dimensional descriptors of molecules; S300: Based on an open-source experimental dataset, a machine learning model is trained by the correlation between one-dimensional, two-dimensional, and three-dimensional molecular descriptors and experimental solvation free energy to obtain a prediction model; S400: Input the molecular descriptors of solute molecules and solvent molecules into the prediction model to obtain the prediction result of the solvation free energy of the solute molecules; In step S200, the one-dimensional descriptor includes molecular weight and number of bonds; the two-dimensional descriptor includes molecular graph features, topological polar surface area (TPSA), and molecular connectivity index; and the three-dimensional descriptor includes molecular surface area and molecular volume. The descriptors of the molecules are APFP, ECFP6, TOPOL, MolProps, and their connection combinations MolPropsAPFP and MolPropsECFP6. In step S200, the missing values ​​in the SMILES transformation are estimated and filled using the K-nearest neighbor classification algorithm to maintain the integrity of the datasets of one-dimensional, two-dimensional, and three-dimensional descriptors; the datasets of one-dimensional, two-dimensional, and three-dimensional descriptors are also normalized to create a consistent scale for the entire dataset. In step S300, the machine learning model is trained using two strategies: a first training strategy and a second training strategy. The first training strategy is based on the offset of the experimental value directly predicted by molecular features, while the second training strategy is based on the difference between the predicted experimental value and the simulated value. The expression for the first training strategy is: Where A represents any sample in the training set. To calculate the free energy obtained using the general Amber force field simulation, These are experimental values ​​obtained from the machine learning model. This represents the prediction offset value corresponding to the first training strategy; The expression for the second training strategy is: For each training set defined by its descriptor in the first training strategy, a machine learning model is applied using 5x cross-validation, resulting in each individual model in the total N of N=5 trained models predicting its own offset. This is the arithmetic mean of these predicted offset values.

2. The method for predicting molecular solvation free energy based on machine learning according to claim 1, characterized in that, In step S300, the open-source experimental data comes from the FreeSolv database. The molecules in the FreeSolv database include a training set and a test set. The training set is used to train the machine learning model, and the test set is used to verify the prediction model.

3. The method for predicting molecular solvation free energy based on machine learning according to claim 1, characterized in that, In step S300, the machine learning models include Support Vector Machine (SVM), Random Forest (RF), Deep Neural Network (DNN), Multiple Linear Regression (MLR), and Extreme Gradient Boosting (XGB), all of which use Bayesian optimization methods to identify the optimal hyperparameter configuration for each model.

4. The method for predicting molecular solvation free energy based on machine learning according to claim 1, characterized in that, In step S300, during the training of the machine learning model, ensemble methods and dimensionality reduction methods are also used for model training. In the ensemble method, the prediction result of the prediction model is obtained by combining the predicted outputs of each machine learning model on average. In the dimensionality reduction method, principal component analysis (PCA) is used to reduce the dimensionality of features in order to improve the model training efficiency.

5. The method for predicting molecular solvation free energy based on machine learning according to claim 1, characterized in that, The solvent molecules are polar solvents, nonpolar solvents, or ionic solvents, and the solute molecules are small organic molecules.

6. The method for predicting molecular solvation free energy based on machine learning according to claim 1, characterized in that, The solvent is water, and the solute is an alcohol, ester, ketone, or aromatic compound.

Images