Asymmetric catalytic reaction prediction method and device based on deep learning

Abstract

The embodiment of the invention discloses an asymmetric catalytic reaction prediction method and device based on deep learning, and relates to the technical field of computers, and the method comprises the steps: obtaining the molecular structure information of a substrate and a catalyst in a to-be-predicted asymmetric catalytic reaction; processing the molecular structure information by using a pre-configured deep learning model to obtain a transition state free energy difference corresponding to the to-be-predicted asymmetric catalytic reaction; the transition state free energy difference is used for representing the stereoselectivity and the absolute configuration corresponding to the asymmetric catalytic reaction to be predicted, so that the stereoselectivity corresponding to the asymmetric catalytic reaction and the absolute configuration of a main product in the reaction can be predicted at the same time, and effective support is provided for rational design of a substrate and a catalyst in the asymmetric catalytic reaction.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method and apparatus for predicting asymmetric catalytic reactions based on deep learning. Background Technology

[0002] In recent years, machine learning and artificial intelligence technologies have been widely used to predict the stereoselectivity of chemical reactions to accelerate catalyst optimization and reaction discovery. However, machine learning models in these technologies have significant limitations when dealing with complex asymmetric catalytic reactions. They struggle to simultaneously predict the stereoselectivity of asymmetric catalytic reactions and the absolute configuration of the major products, thus failing to provide effective support for the rational design of substrates and catalysts. Summary of the Invention

[0003] In view of this, one of the objectives of this application is to provide a method and apparatus for predicting asymmetric catalytic reactions based on deep learning, which can simultaneously predict the stereoselectivity of asymmetric catalytic reactions and the absolute configuration of the major products in the reaction, but cannot provide effective support for the rational design of substrates and catalysts.

[0004] To achieve the above objectives, the technical solution of this application is implemented as follows: In a first aspect, embodiments of this application provide a method for predicting asymmetric catalytic reactions based on deep learning, the method comprising: To obtain molecular structure information of the substrate and catalyst in the asymmetric catalytic reaction to be predicted; The molecular structure information is processed using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted; the transition state free energy difference is used to characterize the stereoselectivity and absolute configuration of the asymmetric catalytic reaction to be predicted. The pre-configured deep learning model is trained based on a training sample set, which includes multiple training samples. Each training sample includes training data and its corresponding training label. The training data includes historical molecular structure information of historical substrates and historical catalysts, and the training label includes the historical transition state free energy difference.

[0005] In one possible implementation, before processing the molecular structure information using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted, the method further includes: Obtain historical molecular structure information of historical reactants and products; Based on historical molecular structure information, determine the changes in the three-dimensional configuration of the chiral center on the historical substrate before and after the reaction; Based on the information on the changes in the three-dimensional configuration of the chiral center, the interaction mode between the historical substrate and the historical catalyst can be determined; Determine the free energy difference of the historical transition state based on the interaction mode; Training samples are obtained by correlating the historical molecular structure information of historical substrates and historical catalysts with the historical transition state free energy difference.

[0006] In one possible implementation, determining the historical transition state free energy difference based on the interaction mode includes: Determine the interaction mode labels corresponding to the interaction modes, and the initial transition state free energy difference in the reaction; The initial transition state free energy difference is adjusted according to the interaction mode label to obtain the historical transition state free energy difference; Specifically, when the interaction mode label is the first label, the initial transition state free energy difference is determined as the historical transition state free energy difference; when the interaction mode label is the second label, the initial transition state free energy difference is inverted and determined as the historical transition state free energy difference.

[0007] In one possible implementation, the pre-configured deep learning model includes a molecular part information module, a reaction center information module, and a pattern interaction module connected in sequence. Molecular structure information is processed using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted, including: Local reaction features are obtained by extracting local features from molecular structure information using a molecular partial information module. The reaction center information module is used to process local reaction features in order to capture global interactions in molecular structure information and obtain global reaction features; By using the pattern interaction module to fuse and filter local and global reaction features, the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted is obtained.

[0008] In one possible implementation, the molecular partial information module includes a word vector embedding unit, a feature extraction unit, and a normalization processing unit connected in sequence. The molecular partial information module is used to extract local features from molecular structure information to obtain local reaction features, including: Molecular structural information is converted into multidimensional character embedding vectors using word vector embedding units; The feature extraction unit is used to extract features from the multidimensional character embedding vector to obtain preliminary features; The preliminary features are standardized using a normalization processing unit to obtain local response features.

[0009] In one possible implementation, the reaction center information module includes a location encoding unit and a multi-head self-attention mechanism unit connected in sequence. The reaction center information module is used to process local reaction features to capture global interactions within molecular structure information, resulting in global reaction features, including: Location-aware features are obtained by using location encoding units to encode local response features. A multi-head self-attention mechanism unit is used to process position-aware features to obtain global dependencies; Global response features are determined based on location-aware characteristics and global dependencies.

[0010] In one possible implementation, the pattern interaction module includes a pooling layer and a fully connected layer connected in sequence; The pattern interaction module is used to fuse and filter local and global reaction features to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted, including: The feature dimensions of local and global response features are concatenated to obtain the concatenated result; The splicing results are downsampled using a pooling layer to obtain the sampling results; By using a fully connected layer to perform residual connection matching on the sampling results, the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted is obtained.

[0011] Secondly, embodiments of this application provide a deep learning-based asymmetric catalytic reaction prediction device, the device comprising: The acquisition module is used to acquire molecular structure information of the substrate and catalyst in the asymmetric catalytic reaction to be predicted; The processing module is used to process molecular structure information using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted; the transition state free energy difference is used to characterize the stereoselectivity and absolute configuration of the asymmetric catalytic reaction to be predicted. The pre-configured deep learning model is trained based on a training sample set, which includes multiple training samples. Each training sample includes training data and its corresponding training label. The training data includes historical molecular structure information of historical substrates and historical catalysts, and the training label includes the historical transition state free energy difference.

[0012] Thirdly, embodiments of this application provide an electronic device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, it implements the method provided in the first aspect.

[0013] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by one or more processors, implements the method provided in the first aspect.

[0014] Fifthly, embodiments of this application provide a computer program product, which includes a computer program that, when executed by one or more processors, implements the method provided in the first aspect.

[0015] The deep learning-based method for predicting asymmetric catalytic reactions provided in this application can acquire the molecular structure information of the substrate and catalyst in the asymmetric catalytic reaction to be predicted, and process the molecular structure information using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted. The transition state free energy difference output by the pre-configured deep learning model in this application can be used to characterize the stereoselectivity and absolute configuration of the asymmetric catalytic reaction to be predicted, thereby achieving simultaneous prediction of the stereoselectivity of the asymmetric catalytic reaction and the absolute configuration of the main products in the reaction, providing effective support for the rational design of substrates and catalysts in asymmetric catalytic reactions. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. It should be understood that the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] The attached diagram is described below: Figure 1 A schematic flowchart illustrating a deep learning-based method for predicting asymmetric catalytic reactions provided in this application embodiment; Figure 2 A schematic diagram of a model provided for an embodiment of this application; Figure 3 This is a schematic diagram illustrating a comparison of model prediction capabilities provided in an embodiment of this application. Figure 4 This is a schematic diagram of an ablation experiment result provided in an embodiment of this application; Figure 5 A schematic diagram illustrating the enantioselectivity of an addition reaction provided in an embodiment of this application; Figure 6 A schematic diagram illustrating the chemical interpretability of a model provided in an embodiment of this application; Figure 7 This is a schematic diagram of the chemical reaction process corresponding to an organocatalytic conjugate addition reaction provided in an embodiment of this application; Figure 8 A schematic diagram illustrating the predicted results of an organocatalytic conjugate addition reaction provided for an embodiment of this application; Figure 9This application provides a schematic diagram of a chemical reaction process corresponding to a photo-oxidation-reduction catalytic asymmetric reaction. Figure 10 A schematic diagram illustrating the prediction results of a photo-redox catalytic asymmetric reaction provided in an embodiment of this application; Figure 11 This is a schematic diagram of the chemical reaction process corresponding to an organocatalytic enamine reaction provided in an embodiment of this application; Figure 12 A schematic diagram illustrating the predicted results of an organocatalytic enamine reaction provided in an embodiment of this application; Figure 13 This application provides a schematic diagram of an interaction mode annotation. Figure 14 A schematic diagram of a reaction result provided in an embodiment of this application; Figure 15 A schematic diagram illustrating the prediction of transition state free energy difference provided in an embodiment of this application; Figure 16 A schematic diagram of the functional modules of an asymmetric catalytic reaction prediction device based on deep learning provided in an embodiment of this application; Figure 17 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application.

[0018] Explanation of reference numerals in the attached figures: 1600. A deep learning-based device for predicting asymmetric catalytic reactions; 1610. Acquisition Module; 1620. Processing module; 1701, Processor; 1702. Memory; 1703. Communication interface; 1710. Bus. Detailed Implementation

[0019] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0020] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0021] To address the technical problems in the background art, embodiments of this application provide a deep learning-based method for predicting asymmetric catalytic reactions, a deep learning-based device for predicting asymmetric catalytic reactions, an electronic device, a computer-readable storage medium, and a computer program product. The deep learning-based method for predicting asymmetric catalytic reactions provided in the embodiments of this application will be described first below.

[0022] Please see Figure 1 , Figure 1 This is a flowchart illustrating a deep learning-based method for predicting asymmetric catalytic reactions, which can be applied to the deep learning-based asymmetric catalytic reaction prediction apparatus or electronic device described in the following embodiments.

[0023] The aforementioned electronic devices include personal computers, servers, mobile devices, cloud computing platforms, and supercomputers. The following section will introduce the deep learning-based asymmetric catalytic reaction prediction method from the perspective of its application in electronic devices. The method specifically includes the following steps 110 to 120.

[0024] Step 110: Obtain the molecular structure information of the substrate and catalyst in the asymmetric catalytic reaction to be predicted.

[0025] The embodiments of this application are applicable to predicting the reaction results of asymmetric catalytic reactions.

[0026] This application does not specifically limit the method of obtaining the above-mentioned molecular structure information. For example, electronic devices can obtain molecular structure information through manual input by the user. As another example, electronic devices can obtain molecular structure information from a molecular structure information database.

[0027] In some embodiments, the aforementioned molecular structure information may be a two-dimensional molecular structure image. A two-dimensional molecular structure image can intuitively display the connection method and spatial layout of atoms in a molecule, enabling electronic devices to quickly identify the basic skeleton of a molecule and the positional relationship of functional groups.

[0028] In some embodiments, the molecular structure information described above may be a SMILES string that can be directly read and processed by an electronic device.

[0029] It should be noted that in asymmetric catalytic reactions, even subtle differences in the molecular structures of the substrate and catalyst can lead to drastically different reaction results. Changes in the position and configuration of specific functional groups in certain substrate molecules, or structural adjustments to the active site of the catalyst, can significantly affect the reaction outcome. Therefore, in the embodiments of this application, obtaining accurate molecular structure information of the substrate and catalyst is beneficial for obtaining an accurate and reliable transition state free energy difference in subsequent step 120, thereby improving the accuracy and reliability of the stereoselectivity and absolute configuration corresponding to the asymmetric catalytic reaction to be predicted.

[0030] Step 120: The molecular structure information is processed using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted; the transition state free energy difference is used to characterize the stereoselectivity and absolute configuration of the asymmetric catalytic reaction to be predicted. The pre-configured deep learning model is trained based on a training sample set, which includes multiple training samples. Each training sample includes training data and its corresponding training label. The training data includes historical molecular structure information of historical substrates and historical catalysts, and the training label includes the historical transition state free energy difference.

[0031] In this embodiment, the electronic device can input the molecular structure information obtained in step 110 into a pre-configured deep learning model, and the pre-configured deep learning model can directly output the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted. It should be noted that the transition state free energy difference in this embodiment can be used simultaneously to characterize the stereoselectivity and absolute configuration of the asymmetric reaction to be predicted.

[0032] The above transition state free energy difference (available) The stereoselectivity of the reaction is closely related to the expression (represented by the expression). In the embodiments of this application, The size can be used to characterize the degree of stereoselectivity. The sign of the slash can be used to characterize different absolute configurations corresponding to the asymmetric reaction to be predicted, specifically to characterize the absolute configuration of the major product in the asymmetric reaction to be predicted. The major product can be understood as the product with the highest proportion of the reaction in the asymmetric catalytic reaction to be predicted, and the absolute configuration of the major product is also the absolute configuration of the product with the highest proportion of the reaction in the asymmetric catalytic reaction to be predicted.

[0033] This application does not impose a specific limit on the number of training samples; the number can be selected according to actual needs.

[0034] The deep learning-based method for predicting asymmetric catalytic reactions provided in this application can acquire the molecular structure information of the substrate and catalyst in the asymmetric catalytic reaction to be predicted, and process the molecular structure information using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted. The transition state free energy difference output by the pre-configured deep learning model in this application can be used to characterize the stereoselectivity and absolute configuration of the asymmetric catalytic reaction to be predicted, thereby achieving simultaneous prediction of the stereoselectivity of the asymmetric catalytic reaction and the absolute configuration of the main products in the reaction, providing effective support for the rational design of substrates and catalysts in asymmetric catalytic reactions.

[0035] It should be noted that the pre-configured deep learning model in the embodiments of this application can be specifically used to predict any one of the following asymmetric catalytic reactions: asymmetric hydrogenation reaction (such as olefin asymmetric hydrogenation reaction), organocatalytic conjugate addition reaction, photoredox catalytic asymmetric reaction, and organocatalytic enamine reaction.

[0036] In one possible implementation, before processing the molecular structure information using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted, the method may include, but is not limited to, steps 210 to 250.

[0037] Step 210: Obtain historical molecular structure information of historical reactants and historical products; Step 220: Based on historical molecular structure information, determine the changes in the three-dimensional configuration of the chiral center on the historical substrate before and after the reaction; Step 230: Determine the interaction mode between historical substrates and historical catalysts based on the information on the changes in the three-dimensional configuration of the chiral center; Step 240: Determine the free energy difference of the historical transition states based on the interaction mode; Step 250: The historical molecular structure information of historical substrates and historical catalysts is correlated with the historical transition state free energy difference to obtain training samples.

[0038] In the embodiments of this application, by acquiring the molecular structure information of historical reactants and products, analyzing the three-dimensional configuration changes of the chiral center, determining the interaction mode of the substrate and catalyst and the transition state free energy difference, and associating this information to form training samples, the model can capture the key factors affecting asymmetric catalytic reactions during the training process, thereby outputting a more accurate transition state free energy difference when predicting new reactions in the future, and achieving reliable prediction of the stereoselectivity and absolute configuration of the reaction.

[0039] The aforementioned historical molecular structure information can be found in the description of molecular structure information in the foregoing embodiments, and will not be repeated here.

[0040] Regarding step 210 above, in some embodiments, the electronic device may use cheminformatics software or professional three-dimensional molecular visualization tools to load three-dimensional structure files of historical reactants and historical products.

[0041] Regarding step 220 above, it should be noted that the chiral center is a carbon atom connecting four different atoms or groups. Electronic devices can determine configurational changes (i.e., changes in the three-dimensional configuration of the chiral center) by comparing the changes in the spatial arrangement of the groups connected to the chiral center before and after the reaction. In the asymmetric catalytic reaction of alkenes, the four groups of the substrate's chiral center have a specific spatial arrangement before the reaction; if the arrangement of the groups changes after the reaction, a configurational change will occur.

[0042] Regarding step 230 above, the electronic device can determine the interaction mode between the substrate and the catalyst during the reaction process based on the direction and characteristics of the change in the three-dimensional configuration of the chiral center. If the configuration change indicates that the interaction between the substrate and the catalyst in a certain direction has led to the configuration change, then the corresponding interaction mode can be determined based on the interaction direction that caused the configuration change.

[0043] Regarding step 240 above, in one possible implementation, determining the historical transition state free energy difference based on the interaction mode includes: Determine the interaction mode labels corresponding to the interaction modes, and the initial transition state free energy difference in the reaction; The initial transition state free energy difference is adjusted according to the interaction mode label to obtain the historical transition state free energy difference; Specifically, when the interaction mode label is the first label, the initial transition state free energy difference is determined as the historical transition state free energy difference; when the interaction mode label is the second label, the initial transition state free energy difference is inverted and determined as the historical transition state free energy difference.

[0044] Specifically, the electronic device can first determine the interaction mode label corresponding to the interaction mode and the initial transition state free energy difference in the reaction (i.e., The initial value). In this embodiment of the application, the training samples are re-labeled. The interaction mode label may include a first label and a second label. When the interaction mode label is the first label (such as "1"), the initial transition state free energy difference is determined as the historical transition state free energy difference (the value and sign do not change); when the interaction mode label is the second label (such as "0"), the initial transition state free energy difference is inverted and determined as the historical transition state free energy difference (the value does not change, but the sign changes).

[0045] Regarding step 250 above, the electronic device can correlate the acquired historical molecular structure information with the determined historical transition state free energy difference to form training samples. The molecular structure information of historical substrates and historical catalysts in a certain historical reaction (such as a specific SMILES string representation) is combined with the historical transition state free energy difference determined in the above steps to serve as multiple training samples for training the deep learning model.

[0046] In one possible implementation, the pre-configured deep learning model includes a molecular part information module, a reaction center information module, and a pattern interaction module connected in sequence. The molecular structure information is processed using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted, including but not limited to steps 310 to 330.

[0047] Step 310: Use the molecular part information module to extract local features of molecular structure information to obtain local reaction features; Step 320: The local reaction features are processed using the reaction center information module to capture the global interactions in the molecular structure information and obtain the global reaction features; Step 330: The local reaction features and global reaction features are fused and screened using the mode interaction module to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted.

[0048] In the embodiments of this application, a three-layer cascaded architecture, namely a molecular part information module, a reaction center information module, and a mode interaction module connected in sequence, can be used to capture from atomic-level local features to intramolecular global interactions, and finally, through information fusion and screening, the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted can be accurately obtained.

[0049] Please see Figure 2 , Figure 2 This is a schematic diagram of a model provided for an embodiment of this application, wherein ChemAHNet represents a pre-configured deep learning model, MoIM represents a molecular part information module, RCIM represents a reaction center information module, and MIM represents a mode interaction module. Figure 2In the middle, the output of ChemAHNet The positive and negative can be further transformed into the corresponding absolute configuration. Figure 2 Not shown in China The specific value.

[0050] Molecular structure contains numerous atoms and chemical bonds. The aforementioned local feature extraction allows us to focus on the specific characteristics of each atom or group of atoms within a molecule, such as atom type, charge distribution, and bond type. By extracting these local features, we can obtain local reaction characteristics, which reflect information such as the local chemical properties and reactivity of the molecule.

[0051] Molecular chemical reactions depend not only on local features, but also on long-range interactions between atoms and groups. The aforementioned reaction center information module can integrate and analyze local features, considering the relationships between various parts of the molecule, thereby obtaining global reaction characteristics. For example, it can analyze the influence of global factors such as electron cloud interactions between different atomic groups and steric hindrance on the reaction.

[0052] In some embodiments, the reaction center information module can process local reaction features through specific algorithms or neural network structures.

[0053] The aforementioned model interaction module can fuse and filter local and global reaction features. Fusion comprehensively considers both local and global molecular information, allowing the model to fully understand the relationship between molecular structure and the reaction. Filtering removes irrelevant or interfering features, retaining only the most useful information for predicting the transition state free energy difference. Through the processing of the model interaction module, the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted can be obtained, providing a valid and reliable theoretical basis for studying the mechanism and reaction conditions of asymmetric catalytic reactions.

[0054] In one possible implementation, the molecular information module includes a word vector embedding unit, a feature extraction unit, and a normalization processing unit connected in sequence; step 310 may include: Molecular structural information is converted into multidimensional character embedding vectors using word vector embedding units; The feature extraction unit is used to extract features from the multidimensional character embedding vector to obtain preliminary features; The preliminary features are standardized using a normalization processing unit to obtain local response features.

[0055] The word vector embedding unit described above converts molecular structure information into multidimensional character embedding vectors. This can be understood as the word vector embedding unit converting the two-dimensional molecular structure image in the aforementioned embodiment into the SMILES string.

[0056] For example, the word vector embedding unit can employ embedding techniques from natural language processing to map basic units in molecular structures, such as atoms, chemical bonds, or functional groups, into multidimensional character embedding vectors.

[0057] The aforementioned feature extraction unit can employ a local perceptual structure of convolutional neural networks, recurrent neural networks, or graph neural networks to perform local correlation analysis on multidimensional character embedding vectors and extract local chemical features of molecules, such as functional group combinations around atoms.

[0058] The normalization processing unit mentioned above can use the Z-Score standardization method to uniformly map the initial features to the same scale, so as to eliminate model bias caused by differences in feature dimension scale. For example, it can avoid bond energy features being over-focused by the model due to their large values, and atomic number features being ignored due to their small values.

[0059] In one possible implementation, the reaction center information module includes a position encoding unit and a multi-head self-attention mechanism unit connected in sequence; step 320 above may include: Location-aware features are obtained by using location encoding units to encode local response features. A multi-head self-attention mechanism unit is used to process position-aware features to obtain global dependencies; Global response features are determined based on location-aware characteristics and global dependencies.

[0060] The aforementioned position encoding unit can encode the position of local reaction features without positional attributes, thereby injecting molecular spatial position information and obtaining position-sensing features. This can avoid the phenomenon that the same local features may have different reactivity due to different positions. For example, the same "-OH" group in a molecule may have completely different effects on asymmetric catalytic reactions depending on whether it is located near a double bond or near a benzene ring.

[0061] In some embodiments, the position encoding unit may employ a position encoding algorithm suitable for molecular structure from related technologies to convert the "three-dimensional coordinate information of atoms / functional groups in the molecule" corresponding to local reaction features into numerical vectors, and fuse them with the local reaction features to obtain position-aware features. For example, the position encoding unit may employ a sine-cosine function position encoding form.

[0062] In a specific implementation, taking the local reaction feature as an example of a 128-dimensional vector, the position encoding unit generates an identical 128-dimensional "position vector" (the value of this position vector is calculated from the X / Y / Z coordinates of the atom in the molecule, for example, through "coordinate normalization + trigonometric function mapping" to ensure that the position vector scale of different molecules is uniform). Then, the "local reaction feature vector" and the "position vector" are fused together by adding or concatenating elements to form a position-aware feature that includes "local chemical properties + spatial position".

[0063] In asymmetric catalytic reactions, the spatial structure of molecules plays a decisive role in the reaction pathway (e.g., the binding of the catalyst and the substrate must satisfy a specific spatial conformation). In this application, the embodiments use position encoding to enable the model to distinguish "local features with the same chemical composition but different positions", which helps to determine accurate and reliable global reaction features.

[0064] In some embodiments, the electronic device may determine global reaction characteristics based on adaptive weighted fusion, according to position-aware features and global dependencies. For example, if the position-aware features focus on "the chemical properties and location of a single locality" and the global dependencies focus on "the interaction strength between different localities", the fused global reaction characteristics can reflect both the basic characteristics of each locality and the synergistic or antagonistic effects between localities. For instance, a substituent in the substrate may weaken the binding ability of the catalyst to the chiral center through global interactions.

[0065] In one possible implementation, the pattern interaction module includes a pooling layer and a fully connected layer connected in sequence; step 330 above may include: The feature dimensions of local and global response features are concatenated to obtain the concatenated result; The splicing results are downsampled using a pooling layer to obtain the sampling results; By using a fully connected layer to perform residual connection matching on the sampling results, the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted is obtained.

[0066] In this embodiment, by splicing, pooling, and fully connected layers to process local and global reaction characteristics, the free energy difference of the transition state of the asymmetric catalytic reaction can be determined. This can effectively integrate and screen key information, improve prediction accuracy, and provide a precise basis for substrate and catalyst design.

[0067] In some embodiments, the activation function in the pre-configured deep learning model includes either the ReLU activation function or the GELU activation function.

[0068] In some embodiments, the output layer of a pre-configured deep learning model includes two parallel branches: a classification branch that uses a Softmax function to output binary classification probabilities, and a regression branch that uses linear activation to output the transition state free energy difference. .

[0069] In some embodiments, the electronic device may employ the AdamW optimizer to optimize a pre-configured deep learning model.

[0070] To test the predictive capabilities of the pre-configured deep learning model ChemAHNet in the above embodiments compared to other state-of-the-art (SOTA) models, an olefin asymmetric hydrogenation dataset (divided into training, validation, and test sets) was used. The test results can be found in [link to test results]. Figure 3 , Figure 3 This is a schematic diagram illustrating a comparison of model prediction capabilities provided in an embodiment of this application.

[0071] It can be observed that, Figure 3 In this application, the pre-configured deep learning model ChemAHNet, provided in the embodiments of this application, achieved an accuracy of up to 88.9% among all models tested. This indicates that the pre-configured deep learning model ChemAHNet can effectively predict the absolute configuration of the major enantiomer (i.e., the major product) by identifying the interaction mode between the substrate and the catalyst, thereby improving the model's understanding of asymmetric catalytic reactions of olefins.

[0072] Furthermore, to verify the rationality and necessity of the three modules in the pre-configured deep learning model ChemAHNet, ablation experiments were conducted. The experimental results can be found in [link to results]. Figure 4 , Figure 4 This is a schematic diagram of an ablation experiment result provided in an embodiment of this application.

[0073] It can be observed that, Figure 4 In the ablation experiment, the prediction accuracy of the model declined after removing any module, indicating that the design of the molecular part information module, reaction center information module and mode interaction module in the above embodiments is reasonable and necessary, and each module has a significant impact on the overall performance of ChemAHNet.

[0074] Furthermore, this application also demonstrates the scalability of the pre-configured deep learning model ChemAHNet by using a dataset of chiral phosphoric acid-catalyzed thiols to N-imides addition reactions. Please see [link to relevant documentation]. Figure 5 , Figure 5This diagram illustrates the enantioselectivity of an addition reaction provided in an embodiment of this application. It can be observed that the pre-configured deep learning model ChemAHNet in this embodiment is not only applicable to asymmetric hydrogenation of olefins, but also has excellent predictive ability for the enantioselectivity of the addition reaction of thiols to N-imides, indicating that ChemAHNet can be extended to other asymmetric catalysis fields.

[0075] In addition, the pre-configured deep learning model ChemAHNet can provide an explanation of spatial and electronic interactions at the atomic level, thus offering a more reliable predictive tool for goal-oriented molecular engineering. For details, please refer to [link to relevant documentation]. Figure 6 , Figure 6 This is a schematic diagram illustrating the chemical interpretability of a model provided in an embodiment of this application.

[0076] It should be noted that, to further illustrate that the pre-configured deep learning model ChemAHNet in this application embodiment possesses good scalability and transferability, and can be applied to the prediction of various other types of asymmetric catalytic reactions, such as the aforementioned organocatalytic conjugate addition reactions, photoredox catalytic asymmetric reactions, and organocatalytic enamine reactions, the following prediction experiments were conducted based on the pre-configured deep learning model ChemAHNet in this application embodiment.

[0077] Prediction Experiment 1: Prediction of conjugate addition reactions in organocatalysis.

[0078] Please see Figure 7 and Figure 8 , Figure 7 This is a schematic diagram of the chemical reaction process corresponding to an organocatalytic conjugate addition reaction provided in an embodiment of this application. Figure 8 This is a schematic diagram illustrating the predicted results of an organocatalytic conjugate addition reaction provided in an embodiment of this application.

[0079] Specifically, the first prediction experiment is based on a dataset containing 90 organocatalytic conjugate addition reactions, covering different types such as Oxa-Michael, Sulfa-Michael, and Phospha-Michael, involving oxygen, sulfur, and phosphorus nucleophiles, and using two types of representative organic catalysts (quinine-based hydrogen bonding catalysts and pyrrolidine-based imine ion catalysts).

[0080] Based on the reaction, ten-fold cross-validation was performed to obtain... Figure 8 The prediction results show that the coefficient of determination in the 10-fold cross-validation results is... The accuracy reached 0.845, and the root mean square error (RMSE) was 0.340 kcal / mol. This indicates that ChemAHNet has good prediction performance in this type of reaction, demonstrating that it can effectively handle the stereoselectivity prediction problem in non-metallic catalytic systems.

[0081] Prediction Experiment 2: Prediction of photo-oxidation-reduction catalyzed asymmetric reactions.

[0082] Please see Figure 9 and Figure 10 , Figure 9 This is a schematic diagram of the chemical reaction process corresponding to a photo-redox catalytic asymmetric reaction provided in an embodiment of this application. Figure 10 This is a schematic diagram illustrating the prediction results of a photo-oxidation-reduction catalytic asymmetric reaction provided in an embodiment of this application.

[0083] Specifically, the second predictive experiment is based on a dataset containing 76 photoredox catalytic reactions, including various reaction types such as asymmetric α-alkylation, radical conjugated addition, cyclization, and deuteration of aldehydes, involving mechanisms such as single-electron transfer (SET) and hydrogen atom transfer (HAT), with chiral control achieved by non-metallic chiral co-catalysts (such as secondary amines, imidazolidinones, or chiral phosphoric acids). Figure 10 ChemAHNet's 10-fold cross-validation results in this type of reaction The value was 0.725 and the RMSE was 0.330 kcal / mol, indicating that ChemAHNet also has good predictive ability for light-driven reactions.

[0084] Predictive Experiment 3: Prediction of organocatalyzed enamine reactions.

[0085] Please see Figure 11 and Figure 12 , Figure 11 This is a schematic diagram of the chemical reaction process corresponding to an organocatalytic enamine reaction provided in an embodiment of this application. Figure 12 This is a schematic diagram illustrating the predicted results of an organocatalytic enamine reaction provided in an embodiment of this application.

[0086] Specifically, Predictive Experiment 3 is based on data from 82 organocatalytic enamine reactions, including classic types such as asymmetric α-fluorination catalyzed by L-proline or chiral imidazolidine ketones, Mannich reactions, and aldehyde-enone additions, covering a variety of electrophilic reagents and typical enamine catalytic mechanisms. For example... Figure 12 ChemAHNet achieves [results] in such reactions. The predicted values ​​of 0.829 and RMSE of 0.385 kcal / mol further confirm its broad applicability to non-hydrogenated chiral induction processes.

[0087] Based on the three cross-domain validation experiments described above, it can be seen that ChemAHNet can achieve high-precision stereoselectivity prediction in various asymmetric catalytic reactions without changing its core architecture, simply by adapting the feature extraction and fine-tuning training to suit the reaction mechanism. This model exhibits excellent transferability and versatility, providing a reliable computational tool for drug synthesis, chiral molecule design, and other fine chemical catalytic development, and possesses broad application prospects.

[0088] In some embodiments, the above-mentioned interaction mode labels can be used to characterize the spatial positional relationship of the key part of the catalyst relative to the reactant plane when the historical substrate (hereinafter referred to as "substrate") and the historical catalyst (hereinafter referred to as "catalyst") interact.

[0089] Taking the asymmetric hydrogenation of olefins as an example, the active sites of the catalyst interact with the carbon-carbon double bonds and surrounding functional groups of the substrate. These interactions exhibit spatial orientation differences. When observing the reaction system from a specific perspective, if the key components of the catalyst (such as groups involved in electron transfer or those affecting the reaction direction due to steric hindrance) are above the plane containing the reactants (mainly olefin substrates containing carbon-carbon double bonds), this sample (i.e., the aforementioned example of a reaction between a group of substrates and the catalyst) can be labeled with the first tag "1"; if it is below the plane containing the reactants, it can be labeled with the second tag "0" (see also [reference needed] for details). Figure 3 (Example of the reaction in the image).

[0090] For details on the marking process, please refer to [link / reference]. Figure 13 , Figure 13 This is a schematic diagram of an interaction mode labeling provided in an embodiment of this application. When the catalyst is located above the plane containing the substrate, the corresponding interaction mode label is the first label "1"; when the catalyst is located below the plane containing the substrate, the corresponding interaction mode label is the second label "0".

[0091] Based on the Cahn-Ingold-Prelog (CIP) rules, which are systematic rules in chemistry used to determine the order of atoms or groups in the stereochemical structure of molecules, it is known that even when the catalyst and substrate are in the same interaction mode, different reaction results may occur due to differences in the priority of the R groups. See [link to relevant documentation] for details. Figure 14 , Figure 14This is a schematic diagram of a reaction result provided in an embodiment of this application, where 1, 2, 3, and 4 represent the priority order of different substituents (4>3>2>1). It can be seen that after the substituents are changed, the reaction result of the object may be either the R configuration or the S configuration. If a sample composed of such a catalyst and substrate is used to train the model, the model cannot correctly distinguish between the R configuration and the S configuration, which will affect the prediction accuracy of the model.

[0092] In this embodiment, the interaction mode between the substrate and the catalyst is labeled with a first label "1" and a second label "0". After obtaining the corresponding interaction mode labels, the corresponding sample can be treated as a whole. During the model training process based on the sample, the changes of different substituents are ignored, thereby improving the prediction accuracy of the model.

[0093] Further, please see Figure 15 , Figure 15 The diagram provided in this application provides a prediction of the transition state free energy difference. It can be seen that without labeling the interaction mode between the substrate and the catalyst, the model trained using samples not labeled in the above embodiments will not be able to predict complex olefin asymmetric hydrogenation reactions with multiple sites (two or more sites).

[0094] It should be noted that the various embodiments described in this application can be combined with each other or implemented individually without conflict, and this application does not limit this.

[0095] Corresponding to the above method embodiments, this application also provides a deep learning-based asymmetric catalytic reaction prediction device. Please refer to [link to relevant documentation]. Figure 16 , Figure 16 A functional module diagram of a deep learning-based asymmetric catalytic reaction prediction device is provided for embodiments of this application. The deep learning-based asymmetric catalytic reaction prediction device 1600 includes: The acquisition module 1610 is used to acquire molecular structure information of the substrate and catalyst in the asymmetric catalytic reaction to be predicted; The processing module 1620 is used to process molecular structure information using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted; the transition state free energy difference is used to characterize the stereoselectivity and absolute configuration of the asymmetric catalytic reaction to be predicted. The pre-configured deep learning model is trained based on a training sample set, which includes multiple training samples. Each training sample includes training data and its corresponding training label. The training data includes historical molecular structure information of historical substrates and historical catalysts, and the training label includes the historical transition state free energy difference.

[0096] The asymmetric catalytic reaction prediction device based on deep learning provided in this application embodiment can achieve, for example: Figure 1 The various processes implemented in the Chinese method embodiments can achieve similar or the same technical effects, and will not be described again here to avoid repetition.

[0097] This application also provides an electronic device; please refer to [link to relevant documentation]. Figure 17 , Figure 17 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application.

[0098] The electronic device may include a processor 1701 and a memory 1702 storing computer program instructions.

[0099] Specifically, the processor 1701 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0100] Memory 1702 may include mass storage for data or instructions. For example, and not limitingly, memory 1702 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 1702 may include removable or non-removable (or fixed) media. Where appropriate, memory 1702 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 1702 is non-volatile solid-state memory.

[0101] In some embodiments, memory 1702 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, generally, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described in the methods provided according to embodiments of this application.

[0102] The processor 1701 implements the method provided in the above embodiments by reading and executing computer program instructions stored in the memory 1702.

[0103] In one example, the electronic device may also include a communication interface 1703 and a bus 1710. The processor 1701, memory 1702, and communication interface 1703 are connected via the bus 1710 and communicate with each other.

[0104] The communication interface 1703 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0105] Bus 1710 includes hardware, software, or both, that couples components of an electronic device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 1710 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0106] Furthermore, in conjunction with the methods provided in the above embodiments, this application embodiment can be implemented using a computer-readable storage medium. This computer-readable storage medium stores computer program instructions; when executed by a processor, these computer program instructions implement any of the methods in the above embodiments.

[0107] Furthermore, in conjunction with the methods provided in the above embodiments, this application embodiment can provide a computer program product to implement the methods. This program product is stored in a storage medium and executed by at least one processor to implement the various processes of the embodiments of the methods provided in the above embodiments, achieving similar or identical technical effects. To avoid repetition, further details are omitted here.

[0108] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0109] The functional blocks shown in the above-described structural diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0110] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0111] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0112] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A deep learning-based method for predicting asymmetric catalytic reactions, characterized in that, The method includes: To obtain molecular structure information of the substrate and catalyst in the asymmetric catalytic reaction to be predicted; The molecular structure information is processed using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted; the transition state free energy difference is used to characterize the stereoselectivity and absolute configuration of the asymmetric catalytic reaction to be predicted. The pre-configured deep learning model is trained based on a training sample set, which includes multiple training samples, each of which includes training data and its corresponding training label. The training data includes historical molecular structure information of historical substrates and historical catalysts, and the training label includes historical transition state free energy differences.

2. The method as described in claim 1, characterized in that, Before processing the molecular structure information using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted, the method further includes: Obtain historical molecular structure information of historical reactants and products; Based on the historical molecular structure information, determine the changes in the three-dimensional configuration of the chiral center on the historical substrate before and after the reaction; Based on the information on the changes in the three-dimensional configuration of the chiral center, the interaction mode between the historical substrate and the historical catalyst is determined; The free energy difference of the historical transition state is determined based on the interaction mode. The training samples are obtained by correlating the historical molecular structure information of the historical substrate and the historical catalyst with the historical transition state free energy difference.

3. The method as described in claim 2, characterized in that, Determining the historical transition state free energy difference based on the interaction mode includes: Determine the interaction mode label corresponding to the interaction mode, and the initial transition state free energy difference in the reaction; The initial transition state free energy difference is adjusted according to the interaction mode label to obtain the historical transition state free energy difference; Specifically, when the interaction mode label is the first label, the initial transition state free energy difference is determined as the historical transition state free energy difference; when the interaction mode label is the second label, the initial transition state free energy difference is inverted and then determined as the historical transition state free energy difference.

4. The method as described in claim 1, characterized in that, The pre-configured deep learning model includes a molecular part information module, a reaction center information module, and a pattern interaction module connected in sequence. The step of processing the molecular structure information using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted includes: The molecular partial information module is used to extract local features from the molecular structure information to obtain local reaction features; The local reaction features are processed using the reaction center information module to capture global interactions in the molecular structure information, thereby obtaining global reaction features. The local reaction features and the global reaction features are fused and filtered using the mode interaction module to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted.

5. The method as described in claim 4, characterized in that, The molecular part information module includes a word vector embedding unit, a feature extraction unit, and a normalization processing unit connected in sequence. The step of using the molecular part information module to extract local features from the molecular structure information to obtain local response features includes: The word vector embedding unit is used to convert the molecular structure information into a multidimensional character embedding vector; The feature extraction unit is used to extract features from the multidimensional character embedding vector to obtain preliminary features; The preliminary features are standardized using the normalization processing unit to obtain the local reaction features.

6. The method as described in claim 4, characterized in that, The reaction center information module includes a location encoding unit and a multi-head self-attention mechanism unit connected in sequence; The local reaction features are processed using the reaction center information module to capture global interactions within the molecular structure information, resulting in global reaction features, including: The location-aware features are obtained by using the location encoding unit to encode the local response features. The location-aware features are processed using the multi-head self-attention mechanism unit to obtain global dependencies; The global response features are determined based on the location-aware features and the global dependencies.

7. The method as described in claim 4, characterized in that, The mode interaction module includes a splicing unit, a pooling layer, and a fully connected layer connected in sequence. The process of fusing and filtering the local and global reaction features using the pattern interaction module to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted includes: The local response features and the global response features are spliced ​​together using the splicing unit to obtain the splicing result. The pooling layer is used to downsample the splicing result to obtain a sampling result; The sampling results are subjected to residual connection matching using the fully connected layer to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted.

8. A deep learning-based device for predicting asymmetric catalytic reactions, characterized in that, The device includes: The acquisition module is used to acquire molecular structure information of the substrate and catalyst in the asymmetric catalytic reaction to be predicted; The processing module is used to process the molecular structure information using a pre-configured deep learning model to obtain the transition state free energy difference corresponding to the asymmetric catalytic reaction to be predicted; the transition state free energy difference is used to characterize the stereoselectivity and absolute configuration corresponding to the asymmetric catalytic reaction to be predicted. The pre-configured deep learning model is trained based on a training sample set, which includes multiple training samples, each of which includes training data and its corresponding training label. The training data includes historical molecular structure information of historical substrates and historical catalysts, and the training label includes historical transition state free energy differences.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by one or more processors, implements the method of any one of claims 1 to 7.

Images