Multi-modal neural network driven reaction site analysis system and method

Abstract

The invention discloses a cross-coupling reaction site analysis system driven by a multi-modal neural network. The cross-coupling reaction site analysis system comprises a data input module, a multi-modal feature extraction module, a multi-modal neural network reasoning module, an analysis result output module and a model iterative optimization module. The system acquires molecular structure, electronic characteristics, reaction environment and historical experimental data, inputs the data into a neural network for fusion reasoning after multi-modal feature coding, and outputs reaction site recognition, activity quantification and side reaction early warning results. Through multi-modal data fusion and cross-modal feature reasoning, accurate recognition and comprehensive risk assessment of reaction sites are realized, the accuracy and reliability of complex molecule analysis are remarkably improved, the dependence on high-cost calculation and experimental trial and error is greatly reduced, and the method is suitable for large-scale popularization and application. And meanwhile, the universality and the self-adaptive optimization capability of the system in different coupling reaction systems are enhanced.

Description

Technical Field

[0001] This invention relates to the field of organic chemical reaction analysis technology, and more specifically, to a multimodal neural network-driven reaction site analysis system and method. Background Technology

[0002] Cross-coupling reactions, as a key means of constructing carbon-carbon and carbon-heteroatom bonds, play a crucial role in fields such as drug synthesis and materials science. The selectivity and efficiency of their reaction pathways highly depend on the precise analysis of multiple potential sites in the reaction substrate, including principal site identification, activity quantification, and prediction of side reaction risks.

[0003] Currently, analytical work in this field faces three major bottlenecks. First, reliance on the subjective judgment of chemical experts makes it difficult to systematically balance electronic effects and steric hindrance for complex molecules, easily overlooking key side reaction pathways and resulting in insufficient prediction accuracy and reliability. Second, single-modal computational simulation methods, such as quantum chemical calculations or molecular fingerprinting, suffer from limitations. While the former has a clear mechanism, it consumes enormous computational resources and is time-consuming, making it difficult to meet the needs of high-throughput research and development. The latter, due to its limited data dimensions, cannot comprehensively reflect the synergistic effects of multiple factors such as molecular structure, electronic properties, and reaction environment, leading to frequent discrepancies between predicted results and experimental phenomena. Third, existing methods generally lack dynamic adaptability, failing to adaptively adjust to different catalyst systems, solvent conditions, and other reaction environments. Each new reaction system requires remodeling, resulting in poor versatility and long development cycles.

[0004] Therefore, there is a need in this field for an intelligent analysis scheme that can deeply integrate multi-dimensional information, has both high precision and high efficiency, and possesses self-learning capabilities, in order to overcome the core pain points of existing technologies such as experience dependence, single dimension, and insufficient adaptability, and provide reliable data-driven decision support for modern organic synthesis. Summary of the Invention

[0005] To address the aforementioned technical problems in related technologies, this invention proposes a multimodal neural network-driven reaction site analysis system and method, which can overcome the above-mentioned shortcomings of the prior art.

[0006] To achieve the above-mentioned technical objectives, the technical solution of the present invention is implemented as follows: A multimodal neural network-driven response site resolution system; This multimodal neural network-driven response site resolution system includes the following modules: The data input module is used to collect multimodal raw data of cross-coupling reactions, including molecular structure data, electronic property data, reaction environment data, and historical experimental data associated with the current reaction. A multimodal feature extraction module is used to perform feature encoding on the original multimodal data to generate multimodal feature vectors; A multimodal neural network inference module is used to fuse and infer the multimodal feature vectors and output the reaction site analysis results, which include reaction site identification, activity quantification and side reaction early warning results; The analysis result output module is used to output the analysis results of the reaction sites; The model iterative optimization module is used to update the parameters of the multimodal neural network inference module based on experimental feedback data.

[0007] Furthermore, the molecular structure data includes two-dimensional structural information and three-dimensional conformational information of the reaction substrate; The electronic property data includes atomic charge, bond order, frontier orbital energy, and Fukui function value; The reaction environment data includes catalyst type, ligand structure, solvent polarity, reaction temperature, and pH value; The historical experimental data includes previous experimental results for the same or similar reaction systems.

[0008] Furthermore, the multimodal feature extraction module includes: The molecular structure feature extraction unit is used to extract features from molecular structure data using graph neural networks and three-dimensional convolutional neural networks. The electronic property feature extraction unit is used to extract features from electronic property data through a fully connected layer and an attention mechanism; The reaction environment feature extraction unit is used to extract features from the reaction environment data through the embedding layer; The historical experiment feature extraction unit is used to extract features from historical experimental data through a temporal convolutional network.

[0009] Furthermore, the multimodal neural network inference module includes: Branch network layers are used to perform preliminary feature mapping on multimodal feature vectors; The cross-modal attention fusion layer is used to dynamically fuse multimodal features through a cross-attention mechanism to generate a global fused feature vector. The inference output layer is used to synchronously output the reaction site identification, activity quantification, and side reaction early warning results based on the global fusion feature vector.

[0010] Furthermore, the inference output layer includes: The reaction site identification branch is used to output the probability that each potential site in the molecule will be the main reaction site. The activity quantification branch is used to output the activity score of each reaction site; The side reaction warning branch is used to output the risk level of side reactions at each non-major site.

[0011] Furthermore, the parsing result output module includes: Site visualization interface, used to annotate reaction sites and their properties on molecular structure diagrams; The data reporting interface is used to output a reaction site selectivity ranking table, detailed activity score data, and side reaction avoidance suggestions. The interface unit is used for data exchange with external software or systems.

[0012] Furthermore, the model iterative optimization module includes: Feedback data acquisition unit is used to collect experimental verification data; The loss calculation unit is used to calculate the error between the predicted value and the experimental true value through a multi-task loss function. The parameter update unit is used to update the parameters of the multimodal neural network inference module using the AdamW algorithm.

[0013] Furthermore, the cross-coupling reaction includes Suzuki coupling, Heck coupling, and Negishi coupling.

[0014] According to another aspect of the present invention, a method for resolving response sites driven by a multimodal neural network is provided; The multimodal neural network-driven response site resolution method includes the following steps: Step 1: Collect multimodal raw data of the cross-coupling reaction. The multimodal raw data includes molecular structure data, electronic property data, reaction environment data, and historical experimental data associated with the current reaction. Step 2: Perform feature encoding on the original multimodal data to generate multimodal feature vectors; Step 3: The multimodal feature vectors are fused and inferred using a multimodal neural network to output the results of reaction site identification, activity quantification, and side reaction early warning. Step 4: Output the results of reaction site identification, activity quantification, and side reaction early warning; Step 5: Update the parameters of the multimodal neural network based on the experimental feedback data.

[0015] Furthermore, the molecular structure data includes two-dimensional structural information and three-dimensional conformational information of the reaction substrate; The electronic property data includes atomic charge, bond order, frontier orbital energy, and Fukui function value; The reaction environment data includes catalyst type, ligand structure, solvent polarity, reaction temperature, and pH value; The historical experimental data includes previous experimental results for the same or similar reaction systems.

[0016] The beneficial effects of this invention are as follows: By integrating molecular structure, electronic properties, reaction environment, and historical experimental data, and utilizing multimodal neural networks for deep fusion and inference of cross-modal features, the system can comprehensively and accurately complete the coordinated output of reaction site identification, activity quantification, and side reaction early warning. This significantly improves the accuracy and reliability of complex molecule analysis, greatly reduces the dependence on high-cost calculations and experimental trial and error, and enhances the system's versatility and adaptability across different cross-coupled reaction systems. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the 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.

[0018] Figure 1 This is an overall architecture diagram of a multimodal neural network-driven reaction site analysis system according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0020] It should be understood that the description of the embodiments of the present invention is only for the purpose of facilitating the description of the embodiments of the present invention and simplifying the description, and is not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the embodiments of the present invention. In addition, the terminology is used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated, unless otherwise expressly and specifically defined.

[0021] like Figure 1 As shown in the embodiment of the present invention, a multimodal neural network driven response site analysis system includes a data input module, a multimodal feature extraction module, a multimodal neural network inference module, an analysis result output module, and a model iteration optimization module. The modules cooperate with each other to complete the accurate analysis of response sites.

[0022] The data input module is responsible for collecting multimodal raw data required for cross-coupling reactions, including molecular structure data (such as the two-dimensional structure SMILES string of the reaction substrate, molecular structure diagram, and three-dimensional conformation file obtained through molecular dynamics simulation), electronic property data (atomic charge, bond order, frontier orbital energy, and Fukui function value obtained through quantum chemical calculation methods such as DFT / B3LYP), reaction environment data (including catalyst type such as Pd(PPh3)4, Ni(dppf)Cl2, ligand structure, solvent polarity such as DMSO, THF, reaction temperature, pH value, and other experimental condition parameters), and historical experimental data associated with the current reaction (including the yield of main products, types of byproducts, and reaction site selectivity data of the same or similar reaction systems).

[0023] The multimodal feature extraction module performs targeted feature encoding on various types of data, transforming them into structured feature vectors. Specifically, the molecular structure feature extraction unit uses graph neural networks (such as GAT graph attention networks) to encode the two-dimensional molecular structure, capturing atomic connections and functional group information, while simultaneously extracting the spatial features of the three-dimensional molecular conformation through a three-dimensional convolutional neural network; the electronic property feature extraction unit normalizes numerical parameters such as atomic charge and bond order through fully connected layers, and uses an attention mechanism to enhance the feature weights of key activity indicators such as the Fukui function value; the reaction environment feature extraction unit uses embedding layers to encode data of categories such as catalysts, ligands, and solvents, while simultaneously standardizing continuous parameters such as temperature and pH; and the historical experiment feature extraction unit uses temporal convolutional networks to extract temporal information such as yield and selectivity from historical experimental data, constructing an empirical feature library for the reaction system.

[0024] The multimodal neural network inference module employs a multi-branch fusion architecture to achieve deep feature fusion and inference. First, the module performs preliminary feature mapping and dimensionality compression on four types of features—molecular structure, electronic properties, reaction environment, and historical experiments—through branch network layers. Then, a cross-modal attention fusion layer dynamically calculates the correlation weights between features from different modalities, achieving deep fusion of multimodal features and generating a global fused feature vector. Finally, the inference output layer simultaneously outputs three types of analytical results: the reaction site identification branch uses a softmax classifier to output the probability of each potential site as a primary reaction site; the activity quantification branch uses a regressor to output the activity score (0-100 points) of each reaction site; and the side reaction warning branch uses a binary classifier to output the side reaction risk level (low risk, medium risk, or high risk) of each non-primary site.

[0025] The analysis results output module presents the analysis results in a visualized and structured format. Its site visualization interface uses different colors to label each reaction site on the molecular structure diagram (e.g., red indicates the primary site, yellow indicates a medium-risk secondary site), and overlays activity scores and risk levels. The data report interface outputs a reaction site selectivity ranking table, detailed activity score data, and side reaction avoidance suggestions. The interface unit provides an API interface to support data exchange and integration with organic synthesis design software such as ChemDraw and experimental data management systems.

[0026] The model iteration and optimization module continuously optimizes the model based on experimental feedback. Its feedback data acquisition unit collects experimental validation data corresponding to the analytical results, including the deviation between the actual reaction main product sites, yields, and predicted results; the loss calculation unit constructs a multi-task loss function to comprehensively calculate classification loss, regression loss, and risk prediction loss; the parameter update unit uses the AdamW algorithm to update the weight parameters of the multimodal neural network based on the loss value through backpropagation, thereby improving the analytical accuracy of the model in specific reaction systems.

[0027] The system's workflow includes the following steps: First, various data of the reaction to be analyzed are collected through the data input module; then, the feature vectors of each modality are generated using the multimodal feature extraction module; next, the feature vectors are input into the multimodal neural network inference module, and after feature fusion, the results of site identification, activity quantification, and side reaction warning are output; then, the analysis results are displayed through the analysis result output module; finally, after obtaining experimental verification, the model parameters are updated through the model iteration optimization module to complete the iterative optimization.

[0028] To facilitate understanding of the above technical solutions of the present invention, the following uses the Suzuki coupling reaction as an example to illustrate the specific application of the multimodal neural network-driven reaction site analysis system described in the present invention.

[0029] In practical application, 4-bromotoluene and phenylboronic acid were used as substrates, Pd(PPh3)4 as catalyst, and DMSO as solvent. Site analysis was performed under reaction conditions of 80℃. First, the substrate's SMILES string, atomic charges and Fukui function values ​​obtained from DFT calculations, catalyst parameters, solvent polarity data, and historical experimental data from 30 similar reactions were imported. Then, the feature extraction module processed the molecular structure, electronic properties, reaction environment, and historical experimental data. During the fusion inference stage, the system identified the coordination association between the Pd catalyst and the para-carbon atom in 4-bromotoluene (which has the highest Fukui function value), outputting that this para-carbon is the main reaction site (probability 98%), with an activity score of 92, and the risk level of the adjacent carbon side reaction is low. Finally, the main and side sites were marked in the visualization interface, and an analysis report was generated suggesting that the reaction temperature be controlled at 75-85℃ to maintain the selectivity of the main site. Experimental verification showed that the main product was p-methylbiphenyl (yield 91%), with a deviation from the predicted result of less than 2%. This feedback data was used for further optimization of the model parameters.

[0030] In summary, the following beneficial effects are achieved by utilizing the above-described technical solution of the present invention: Improved resolution accuracy: The design of multimodal data and cross-attention fusion improves the system’s accuracy in identifying reaction sites of complex molecules (such as multisubstituted heterocyclic compounds) to over 90%, which is far higher than traditional manual judgment (about 65%) and single-modal calculation (about 75%). Reduced computational costs: Compared to pure quantum chemical calculations (which can take hours to perform a single calculation of complex molecules), this system is based on pre-trained neural network inference, and a single analysis takes only 1-5 minutes, improving efficiency by more than 100 times. Reduce experimental trial and error: Through the side reaction early warning function, potential side reaction sites can be predicted in advance, reducing the number of invalid experiments and reducing time and material costs in drug synthesis and material development; Enhanced versatility: Supports various cross-coupling reaction types such as Suzuki, Heck, and Negishi, and can be adapted to new reaction systems (such as novel catalysts and novel substrates) through model iteration, eliminating the need for repeated development of analytical models and making it widely applicable.

[0031] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multimodal neural network-driven response site resolution system, characterized in that, Includes the following modules: The data input module is used to collect multimodal raw data of cross-coupling reactions, including molecular structure data, electronic property data, reaction environment data, and historical experimental data associated with the current reaction. A multimodal feature extraction module is used to perform feature encoding on the original multimodal data to generate multimodal feature vectors; A multimodal neural network inference module is used to fuse and infer the multimodal feature vectors and output the reaction site analysis results, which include reaction site identification, activity quantification and side reaction early warning results; The analysis result output module is used to output the analysis results of the reaction sites; The model iterative optimization module is used to update the parameters of the multimodal neural network inference module based on experimental feedback data.

2. The multimodal neural network-driven response site resolution system according to claim 1, characterized in that, The molecular structure data includes two-dimensional structural information and three-dimensional conformation information of the reaction substrate; The electronic property data includes atomic charge, bond order, frontier orbital energy, and Fukui function value; The reaction environment data includes catalyst type, ligand structure, solvent polarity, reaction temperature, and pH value; The historical experimental data includes previous experimental results for the same or similar reaction systems.

3. The multimodal neural network-driven reaction site resolution system according to claim 1, characterized in that, The multimodal feature extraction module includes: The molecular structure feature extraction unit is used to extract features from molecular structure data using graph neural networks and three-dimensional convolutional neural networks. The electronic property feature extraction unit is used to extract features from electronic property data through a fully connected layer and an attention mechanism; The reaction environment feature extraction unit is used to extract features from the reaction environment data through the embedding layer; The historical experiment feature extraction unit is used to extract features from historical experimental data through a temporal convolutional network.

4. The multimodal neural network-driven response site resolution system according to claim 1, characterized in that, The multimodal neural network inference module includes: Branch network layers are used to perform preliminary feature mapping on multimodal feature vectors; The cross-modal attention fusion layer is used to dynamically fuse multimodal features through a cross-attention mechanism to generate a global fused feature vector. The inference output layer is used to synchronously output the reaction site identification, activity quantification, and side reaction early warning results based on the global fusion feature vector.

5. The multimodal neural network-driven reaction site resolution system according to claim 4, characterized in that, The inference output layer includes: The reaction site identification branch is used to output the probability that each potential site in the molecule will be the main reaction site. The activity quantification branch is used to output the activity score of each reaction site; The side reaction warning branch is used to output the risk level of side reactions at each non-major site.

6. The multimodal neural network-driven reaction site resolution system according to claim 1, characterized in that, The parsing result output module includes: Site visualization interface, used to annotate reaction sites and their properties on molecular structure diagrams; The data reporting interface is used to output a reaction site selectivity ranking table, detailed activity score data, and side reaction avoidance suggestions. The interface unit is used for data exchange with external software or systems.

7. The multimodal neural network-driven response site resolution system according to claim 1, characterized in that, The model iterative optimization module includes: Feedback data acquisition unit is used to collect experimental verification data; The loss calculation unit is used to calculate the error between the predicted value and the experimental true value through a multi-task loss function. The parameter update unit is used to update the parameters of the multimodal neural network inference module using the AdamW algorithm.

8. The multimodal neural network-driven response site resolution system according to claim 1, characterized in that, The cross-coupling reactions include Suzuki coupling, Heck coupling, and Negishi coupling.

9. A multimodal neural network-driven method for resolving response sites, characterized in that, Includes the following steps: Step 1: Collect multimodal raw data of the cross-coupling reaction. The multimodal raw data includes molecular structure data, electronic property data, reaction environment data, and historical experimental data associated with the current reaction. Step 2: Perform feature encoding on the original multimodal data to generate multimodal feature vectors; Step 3: The multimodal feature vectors are fused and inferred using a multimodal neural network to output the results of reaction site identification, activity quantification, and side reaction early warning. Step 4: Output the results of reaction site identification, activity quantification, and side reaction early warning; Step 5: Update the parameters of the multimodal neural network based on the experimental feedback data.

10. The method for resolving response sites driven by a multimodal neural network according to claim 9, characterized in that, The molecular structure data includes two-dimensional structural information and three-dimensional conformation information of the reaction substrate; The electronic property data includes atomic charge, bond order, frontier orbital energy, and Fukui function value; The reaction environment data includes catalyst type, ligand structure, solvent polarity, reaction temperature, and pH value; The historical experimental data includes previous experimental results for the same or similar reaction systems.

Images