Combined prediction method and system for chemical reaction energy barrier and transition state based on bimodal reaction characterization architecture and diffusion model

Abstract

A chemical reaction energy barrier and transition state combined prediction method based on a bimodal reaction characterization architecture and a diffusion model comprises the following steps: generating standard input data containing three-dimensional geometric information and an atomic mapping alignment relationship through an automatic data standardization preprocessing module; a bimodal reaction representation skeleton is constructed, the skeleton enhances perception of three-dimensional geometric information by using an isovariant graph encoder, and global chemical context information is captured by using a self-supervised pre-trained sequence encoder; through a cross-modal alignment fusion module, the two kinds of modal information are efficiently fused to form a unified reaction expression; based on the unified representation, high-precision chemical reaction energy barrier prediction is performed through a downstream regression prediction network, and a transition state atomic coordinate structure is generated by using an isovariant diffusion model. According to the method, the generalization ability of the model on a small-scale transition state data set is effectively improved, and an efficient and accurate prediction tool is provided for computational chemistry and drug design.

Description

Technical Field

[0001] This invention belongs to the fields of computational chemistry, artificial intelligence, and deep learning, specifically relating to a chemical reaction energy barrier (Activation Energy Barrier) based on a bimodal reaction characterization architecture and diffusion model. A method and system for joint prediction of ) and transition state (TS). Background Technology

[0002] The transition state (TS) of a chemical reaction is the highest energy structure connecting reactants and products; the difference between its energy and the energy of the reactants is defined as the reaction energy barrier. , which are the core kinetic parameters that determine the reaction rate. Accurate acquisition of this information is crucial for modern chemical synthesis and materials science.

[0003] Traditionally, accurate prediction of energy barriers and transition state (TS) structures has relied primarily on high-precision quantum chemical (QC) calculations, such as density functional theory (DFT). While QC calculations are accurate and reliable, they are extremely costly, especially for large molecular systems containing hundreds of atoms or in high-throughput virtual screening scenarios, where computation times can reach hours or even days, severely limiting their large-scale application. Furthermore, finding transition states requires complex optimization processes (e.g., the perturbative elastic band method, NEB) and is highly dependent on the quality of the initial structure guess, resulting in low success rates and low levels of automation.

[0004] In recent years, machine learning (ML) models have made progress in accelerating chemical reaction prediction, but they still face the following fundamental challenges in transition state prediction tasks: (1) Heavy dependence on precise geometric information and data sparsity: Energy barriers and transition state structures are essentially three-dimensional problems, and model performance is highly sensitive to precise molecular three-dimensional conformations (3D geometric information). However, high-precision transition state datasets are costly to construct and are usually small in scale, resulting in data sparsity. Existing ML methods suffer from severe generalization ability and robustness when dealing with inaccurate or coarse 3D inputs generated by low-precision methods (such as force field calculations).

[0005] (2) Insufficient reaction characterization capability: Energy barrier prediction is essentially a downstream task of reaction characterization. Many existing ML models only focus on molecular structure and lack effective modeling of reaction-level information, i.e., they cannot simulate the interactions between atoms and the formation or breaking of chemical bonds during the reaction process. Simple feature splicing cannot effectively integrate global chemical context information, resulting in poor generalization ability of the model to complex and unseen chemical reaction types.

[0006] (3) Existing graph neural networks (GNNs) also have limitations. Traditional message-passing-based GNNs (MPNNs) mainly focus on the two-dimensional topological structure of molecules. Although they can effectively capture local features, their ability to perceive and utilize three-dimensional geometric information such as spatial distances between atoms, bond angles, and torsion angles is limited. This makes it difficult for them to accurately characterize the fine spatial configuration changes during the reaction process, especially in tasks such as energy barrier prediction, which have extremely high requirements for geometric accuracy.

[0007] In summary, existing technologies lack a transition state prediction model that can effectively integrate three-dimensional geometric information and global chemical context while exhibiting high robustness to input data quality. This invention aims to overcome these challenges through an innovative solution based on reaction graph / text multimodal pre-training technology and generative model frameworks such as diffusion models. Summary of the Invention

[0008] To address the shortcomings of the existing technologies, this invention provides a method and system for jointly predicting the energy barrier and transition state of chemical reactions based on a dual-modal reaction characterization architecture and a diffusion model. By integrating graph neural networks (GNN), pre-trained sequence models, and isovariant diffusion models, it achieves accurate modeling of energy changes and geometric configurations during chemical reactions, providing an efficient and accurate prediction tool for catalyst design, reaction mechanism exploration, and drug synthesis pathway optimization.

[0009] This invention is achieved through the following technical solutions.

[0010] A method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model is characterized by the following steps: S1. Data Acquisition and Standardization: Input the raw data of the chemical reaction, including SMILES expressions and optional 3D conformation information.

[0011] The automated processing module performs the following tasks: SMILES standardization, data deduplication, atomic mapping relationship solving, and reaction expression validity screening.

[0012] A low-cost two-step method is used to generate molecular 3D coordinates: first, coarse optimization is performed using the MMFF force field of RDKit, and then fine optimization is performed using the xTB program for single-point calculation, balancing computational cost and coordinate quality.

[0013] Perform atomic order remapping on the existing transition state coordinates to ensure consistency with the standard SMILES order.

[0014] S2. Dual-modal feature encoding and fusion: A pre-trained multimodal fusion model is constructed, which includes an isomorphic graph encoder and a pre-trained sequence encoder to extract three-dimensional features and global chemical semantic features from the standardized input, respectively. The two features are then aligned and fused through a cross-modal fusion mechanism to generate a unified chemical reaction representation vector. Among them, the isomorphic map encoder: An explicit alignment relationship between reactants and products is constructed based on atomic mapping, and two independent MPNNs are used to process the reactant and product maps respectively.

[0015] A geometric bias self-attention mechanism is introduced, which encodes the Euclidean distance, bond angle, and torsion angle between atoms as attention correction terms to enhance the perception of three-dimensional geometry.

[0016] A geometric graph neural network structure such as UniMol Layer is used to achieve robust encoding of low-precision 3D input.

[0017] Sequence encoder: A chemical word segmenter is built based on regular expressions to perform lexical segmentation on SMILES expressions, and supports the embedding of reaction separators (">>").

[0018] A pre-trained Chemformer model (based on the BART architecture) is used as the sequence encoding backbone to extract global chemical context semantic information.

[0019] S3. Transition State Energy Barrier Prediction and Structure Generation: Based on the unified chemical reaction representation vector, the following two tasks are performed simultaneously: III. Output the predicted activation energy barrier value of the chemical reaction through a regression prediction network; IV. A three-dimensional atomic coordinate structure of the chemical reaction transition state is generated by using an isovariant diffusion model conditioned on the aforementioned unified representation.

[0020] Furthermore, the data standardization preprocessing module includes SMILES standardization, data deduplication, chemical reaction atom mapping acquisition, expression validity screening, and molecular coordinate acquisition or remapping. Details are as follows: The rdkit-python class loads the text in the form of molecular SMILES or molecular formulas from the dataset, exports and saves the standardized SMILES expressions, and completes the splicing of reactant and product SMILES to obtain the reaction expression RXN, which serves as the standard input for the subsequent model.

[0021] Data deduplication: If multiple energy barriers or other data points appear for the same reaction in the dataset, these data indicators are averaged, and only that indicator is retained for that reaction.

[0022] For all reaction expressions, the open-source tool rxn-mapper is used to solve for the atom-to-atom mapping relationships of reactants / products. This serves two purposes: firstly, it meets the model input requirements; secondly, it clarifies the changes of each atom in the reaction, enabling the model to capture information about the reaction process. After completing the atom mapping, a confidence score is obtained, and reaction expressions with excessively low scores are discarded to avoid interfering with the stability of model training.

[0023] The reaction expressions are screened for validity: the total charge of reactants / products is calculated separately to filter out reactions with uneven charge distribution; reactions that cannot be properly resolved by SMILES are filtered out; and the frequency of quantitative indicators is statistically analyzed to remove reactions corresponding to extremely rare extreme values.

[0024] The molecular coordinates in the final effective reaction equation are calculated using a low-cost two-step method to obtain reasonably high-quality 3D coordinates: First, a rough generation is performed using RDkit-python's MMFFOptimizeMolecule method to generate a coarse 3D molecular structure; then, the xtb-single-point calculation program is used to further refine the coarse structure to obtain relatively accurate spatial coordinates. This avoids excessive reliance on high-precision computational chemistry.

[0025] For the transition state coordinate prediction task, the dataset contains transition state coordinates obtained through computational chemistry, but the atomic order in the coordinate data is inconsistent with the order in the standard SMILES, requiring atomic remapping. The core approach is to use the GetSubstructMatches method provided by RDKit to perform isomorphic matching of the molecular diagram and obtain the matching mapping. That is, in computational chemistry, the position of the target atom in the normalized molecular expression is obtained in the molecular coordinate system order, and after obtaining the matching mapping, the original coordinate array is normalized and reordered.

[0026] The design of the isovariant reaction map encoder is characterized by: (1) building a reaction characterization learning mechanism based on atomic mapping, and (2) using a geometric graph neural network structure to encode three-dimensional conformation data, wherein the GNN structure captures geometric information based on the UniMolLayer skeleton.

[0027] The reaction characterization learning mechanism in the isomorphic graph encoding module is characterized by utilizing atom-mapping to explicitly model the alignment relationship between reactants and products, thereby establishing an information transfer path across the molecular graph and simulating the interactions and evolution between atoms during the reaction process. For reactants and products, two different MPNN networks are used for aggregation and updating: , Corresponding message passing mechanism, Indicates the aggregation method of node messages. This indicates the update method for any node. Finally, through atomic mapping relationships, after characterizing the corresponding atoms in the reactants / products, collaborative learning is performed to complete an additional interaction on the reactants / products to capture reaction-level information. = .

[0028] To address the weakness of traditional GNNs in perceiving 3D spatial information, this invention replaces the traditional message-passing-based GNN (MPNN) encoding layer with a UniMol Layer. Its key feature is the introduction of geometric bias into the Transformer's self-attention mechanism to encode 3D information. This mechanism transforms the geometric relationships between atoms into a correction term for the attention score.

[0029] The geometric deviation self-attention mechanism is characterized by, for any two atoms in the diagram... and The attention calculation during message passing is as follows: ,in, From atoms Passed to the atom The message; and They are atoms and The query and key vector; Self-attention calculation representing the standard; It is a geometric deviation term that explicitly encodes the atoms. and The three-dimensional geometric relationships between atoms are derived from the Euclidean distance between atoms and other geometric quantities. Specifically, to supplement sufficient geometric information to compensate for the low-precision three-dimensional coordinate calculations, the following geometric quantities are additionally calculated as follows: Input , It is a distance encoding function. It's a key corner. It is the torsion angle.

[0030] The construction of the sequence model word segmenter structure includes: segmenting molecular representations in the SMILES language system based on regular expressions, and adding delimiters and other symbols to the regular expression system to support the segmentation of reaction expressions. Chemical reaction expressions are segmented from SMILES expressions into a series of text sequences based on chemical language morphemes. The vocabulary provided by the Chemformer model is expanded by adding reaction delimiters ('>>') and other special symbols, while simultaneously increasing its embedding matrix. For chemical reaction expressions... The sequence is obtained by passing it through an expanded word segmenter and an embedding layer. The atomic sequences from the graph encoder are flattened according to the order in which the SMILES expression appears, resulting in the graph sequence. .

[0031] The graph embedding-text alignment mask mechanism involves designing a cross-attention mechanism mask for aligned tokens to integrate information from the graph embedding space into a text sequence. For the input chemical expression, it first labels different molecules. Simultaneously, for graph sequence nodes and text sequence tokens, molecular tags are assigned according to their respective molecules. or Construct a special mask. Its characteristic is that, for position , If and equal to ,or This is the delimiter. This mask is used in the graph-text modality alignment mechanism to enhance the embedding representation of text sequences.

[0032] The complete bimodal alignment mechanism utilizes cross-attention and residual connections as modal alignment methods. The key to the cross-attention mechanism lies in using a graph embedding-text alignment mask. The attention mask part of the cross-attention mechanism is used, and the SMILES sequence T and the graph sequence G are used as the query vector and key vector parts of the cross-attention mechanism, respectively. The residual connection part is an additive connection using the cross-attention calculation result and the sequence T itself.

[0033] The implementation of the cross-attention mechanism is as follows: ,in .

[0034] The residual connection mechanism consists of the following: Where O represents the calculation result of the cross-attention mechanism.

[0035] Enhanced response The input will be fed into a pre-trained sequence encoder backbone. This sequence encoder employs a Chemformer model based on the BART (Bidirectional Autoregressive Transformer) architecture. The backbone contains a bidirectional encoder and an autoregressive decoder for further extraction of global chemical context and deep semantic information. Specifically, the enhanced sequence... By performing feature interaction and deep extraction through multiple Transformer Encoder layers, the final response latent vector representation containing image-text dual-modality information is obtained. This representation not only aggregates the long-range dependencies of the SMILES text sequence through an attention mechanism, but also incorporates three-dimensional geometric features from the graph encoder, forming a unified reaction representation. This provides high-dimensional feature support for subsequent energy barrier prediction and structure generation.

[0036] The structure used in this invention differs from the standard Chemformer structure. It can be used directly as input to the encoder without going through the standard Chemformer word embedding layer again.

[0037] The model pre-training module includes the following three pre-training objectives: A masked language model strategy is employed to randomly mask some of the tokens of reactants or products, constructing a corrupted input sequence. The goal is to maximize the probability that the model reconstructs the original SMILES sequence, and the pre-training loss function is defined. To predict the cross-entropy loss between the token and the real token: Sequence generation pre-training – forward reaction prediction: This aims to force the model to learn explicit rules for the transformation from reactants to products, a crucial prerequisite for understanding transition states. The reactant SMILES sequence is used for this purpose. As input to the encoder, the decoder is required to autoregressively generate the corresponding product SMILES sequence. .

[0038] Maximizing the conditional probability of the model generating the correct product sequence given the reactant sequence is equivalent to minimizing the following loss: ,in, Is the product sequence in Real Token at all times This is the token generated earlier. This task directly trains the model as a translator of chemical reactions, thereby enhancing its ability to encode reaction coordinates.

[0039] Sequence generation pre-training—chemical reaction conditions: This aims to enable the model to infer the external environmental factors (such as solvents, catalysts, etc.) required to achieve the transformation from the structural changes of the reaction itself, thereby incorporating kinetic information into the reaction characterization. The reaction sequence is then used for pre-training. As input to the encoder, the decoder is required to autoregressively generate SMILES sequences corresponding to the reaction conditions. .

[0040] That is, to maximize the model's ability to generate the correct reaction condition sequence under given reaction sequence conditions. The conditional probability is equivalent to minimizing the following loss: ,in, Is the reaction condition sequence in Real Token at all times These are the tokens generated previously. This task provides important dynamic constraint information for TS prediction.

[0041] The key aspect of the transition state energy barrier prediction step lies in the analysis of unified mode characteristics. Extract the feature vector at a specific location or perform mean pooling on all token outputs to obtain a fixed-length reaction-level feature vector. , to feature vector The input is fed into a downstream regression prediction network. This regression network consists of a multilayer perceptron, containing linear layers and nonlinear activation functions (such as ReLU), and ultimately outputs a scalar value, namely the predicted response energy barrier. Construct the loss function for the regression task. Mean squared error (MSE) is typically used to measure the difference between the predicted energy barrier and the actual energy barrier. The differences between them .

[0042] Automated Evaluation – The energy barrier prediction section performs key evaluations on the current pre-training and energy barrier prediction components. For the pre-training phase, token prediction accuracy and perplexity on the validation set are used as evaluation metrics to monitor the model's mastery of chemical language. In the energy barrier prediction fine-tuning task, the following statistical metrics are used to evaluate the model's predictive performance: Mean squared error (MSE) measures the squared average of the prediction error and reflects the overall goodness of fit of the model.

[0043] Mean Absolute Error (MAE) It measures the average absolute deviation between predicted and actual values ​​and is less sensitive to outliers.

[0044] Root mean square error (RMSE). It is used to measure the deviation between the predicted value and the actual value. The unit is consistent with the energy barrier unit (e.g., kcal / mol), which makes it easy to intuitively evaluate the prediction accuracy.

[0045] The subsequent transition state structure generation module will utilize the established bimodal representation ( This addresses the most challenging problem of transition state structure generation and employs an isovariant diffusion model to predict the atomic coordinates of the transition state. The diffusion model uses a Markov chain to gradually diffuse random noise. Transformed into a molecular structure with physical meaning. .

[0046] To ensure the chemical validity of the generated TS structure in three-dimensional space, the prediction model... An E(3) equivariant graph neural network is used as the backbone. The E(3) equivariance ensures that the physical quantities of atomic coordinates remain unchanged under rotation and translation operations in three-dimensional space, which is a prerequisite for generating a reliable molecular structure.

[0047] By minimizing prediction noise and actual noise The differences between them are used to train the structure generation module: .

[0048] The noise prediction skeleton in the graph neural network of the diffusion model must satisfy the requirement of equivariance, that is: noise prediction skeleton The characteristic is that for any rotation Peaceful relocation The operation must meet the following requirements: This means input coordinates The rotation and translation (on the left) will cause noise / gradient in the output prediction. Rotate by an equal amount (right side).

[0049] The key to implementing the aforementioned conditional structure generation mechanism lies in representing a unified reaction. Injected as conditional information into the denoising process In this process, the structure generation is guided. This ensures that the generated TS structure is not only geometrically reliable, but also optimized based on predicted energy barrier kinetic information.

[0050] The key denoising steps of its diffusion model are: .

[0051] in, It is the first Noise data of the step, It is a parameterized denoising function. This is the dual-mode conditional characteristic after fusion. This conditional generation process can predict the relative coordinates of each atom in the transition state stage, starting from the reactant / product coordinates.

[0052] The generated predicted structures will be evaluated using an automated evaluation module—the structure generation evaluation module. This module assesses the geometric accuracy and chemical plausibility of the transition state (TS) 3D structures generated by the isovariant diffusion model and consists of two parts: Root Mean Square Shift (RMSD) Evaluation: Calculating the predicted transition state atom coordinates Compared with the true transition state coordinates obtained using high-precision quantum chemical calculations The root mean square displacement between them.

[0053] To eliminate the effects of rigid rotation and translation, the Kabsch algorithm must be used to perform optimal alignment of the two sets of coordinates before calculation.

[0054] The formula for calculating RMSD is: ,in, The number of atoms involved in the reaction. Let be a rotation matrix. This is the translation vector. A low RMSD value (typically less than 0.5 Å) indicates a high degree of agreement between the predicted and actual structures.

[0055] Structural validity check: In addition to geometric coordinate errors, validity verification needs to be conducted in conjunction with prior chemical knowledge, including: Bond length and bond angle check: Verify that the length and angle of the key chemical bonds in the generated TS structure are within a reasonable physical range, and exclude non-physical atomic overlap or breakage.

[0056] Imaginary frequency verification: Verify whether the predicted structure contains only one imaginary frequency after Hessian matrix analysis. This is a necessary and sufficient condition for determining that the structure is a first-order saddle point (i.e., a transition state).

[0057] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: This invention provides a method and system for jointly predicting chemical reaction energy barriers and transition states based on isovariant graph networks, pre-trained sequence backbones, and diffusion models. With multimodal fusion technology at its core, it effectively integrates three-dimensional geometric information and global chemical context, overcoming the challenges of existing models' high dependence on precise geometric information and insufficient reaction characterization capabilities. By employing isovariant graph networks such as the UniMol Layer, the robustness and generalization ability of the model when processing low-precision three-dimensional conformational data are significantly improved. Simultaneously, by utilizing Chemformer pre-training and isovariant diffusion models, high-precision energy barrier prediction and high-confidence transition state structure generation are achieved, providing an efficient and accurate prediction tool for key R&D stages such as catalyst design, exploration of novel reaction mechanisms, and optimization of drug molecule synthesis pathways. Attached Figure Description

[0058] Other features, objects, and advantages of the invention will become more apparent from the detailed description of the non-limiting embodiments with reference to the following accompanying drawings: Figure 1 This is a flowchart of a preferred embodiment of the present invention, showing a method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and a diffusion model. Detailed Implementation

[0059] The embodiments of the present invention are described in detail below: These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

[0060] This embodiment provides a method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model. Based on automated data processing and graph / text multimodal fusion algorithms, it utilizes data standardization, atomic mapping, and 3D conformation generation to transform real chemical reaction data into a unified multimodal representation that the model can process. Through deep learning algorithms, this system efficiently fuses the 3D geometric features and 1D semantic information of molecules using an isovariant graph network and a pre-trained sequence backbone, constructing a unified reaction representation with high insight into the chemical reaction process. Based on this, a regression predictor returns accurate reaction energy barrier prediction results, and an E(3) isovariant diffusion model returns a highly reliable 3D atomic coordinate structure of the transition state. This method effectively characterizes the uncertainties in energy changes and geometric configurations in chemical reactions, exhibiting strong robustness, high precision, and physical rationality. This system is beneficial for forming a fundamental technical support for the quantitative evaluation of complex chemical reactions, the generation of transition state structures, and mechanism prediction. It provides strong support for the intelligent design of new catalysts and drug molecules, reaction pathway optimization, and basic chemical research, contributing to the efficient development of cheminformatics and synthetic chemistry.

[0061] This example provides a method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model. Figure 1 As shown, it includes the following steps: Step S1: Collect raw data of chemical reaction, including reaction expression and three-dimensional conformation information, and after automated processing, form standardized input data containing sequence information, geometric information and atomic mapping relationship; Step S2: Construct a pre-trained multimodal fusion model, which includes an isomorphic graph encoder and a pre-trained sequence skeleton. Encode the features of the standardized input data and align them through a cross-modal fusion mechanism to obtain a unified chemical reaction representation. Step S3: Based on the unified chemical reaction representation, the data is input into the downstream regression prediction network to predict the activation barrier index of the chemical reaction. Step S4: Build an automated evaluation-energy barrier prediction module to quantify the model prediction effect of the pre-trained index and the activation energy barrier index.

[0062] Step S5: Based on the unified chemical reaction representation, a three-dimensional atomic coordinate structure of the chemical reaction transition state stage is generated using the isovariant diffusion model. Step S6: Build an automated evaluation module for predicting and assessing transition state atomic coordinates to evaluate the transition state atomic coordinate structure and quantify the model's prediction performance.

[0063] In a preferred embodiment, step S1 includes: The rdkit-python class loads the text in the form of molecular SMILES or molecular formulas from the dataset, exports and saves the standardized SMILES expressions, and completes the splicing of reactant and product SMILES to obtain the reaction expression RXN, which serves as the standard input for the subsequent model.

[0064] Data deduplication: If multiple energy barriers or other data points appear for the same reaction in the dataset, these data indicators are averaged, and only that indicator is retained for that reaction.

[0065] For all reaction expressions, the open-source tool rxn-mapper is used to solve for the atom-to-atom mapping relationships of reactants / products. This serves two purposes: firstly, it meets the model input requirements; secondly, it clarifies the changes of each atom in the reaction, enabling the model to capture information about the reaction process. After completing the atom mapping, a confidence score is obtained, and reaction expressions with excessively low scores are discarded to avoid interfering with the stability of model training.

[0066] The reaction expressions are screened for validity: the total charge of reactants / products is calculated separately to filter out reactions with uneven charge distribution; reactions that cannot be properly resolved by SMILES are filtered out; and the frequency of quantitative indicators is statistically analyzed to remove reactions corresponding to extremely rare extreme values.

[0067] The molecular coordinates in the final effective reaction equation are calculated using a low-cost two-step method to obtain reasonably high-quality 3D coordinates: First, a rough generation is performed using RDkit-python's MMFFOptimizeMolecule method to generate a coarse 3D molecular structure; then, the xtb-single-point calculation program is used to further refine the coarse structure to obtain relatively accurate spatial coordinates. This avoids excessive reliance on high-precision computational chemistry.

[0068] For the transition state coordinate prediction task, the dataset contains transition state coordinates obtained through computational chemistry, but the atomic order in the coordinate data is inconsistent with the order in the standard SMILES, requiring atomic remapping. The core approach is to use the GetSubstructMatches method provided by RDKit to perform isomorphic matching of the molecular diagram and obtain the matching mapping. That is, in computational chemistry, the position of the target atom in the normalized molecular expression is obtained in the molecular coordinate system order, and after obtaining the matching mapping, the original coordinate array is normalized and reordered.

[0069] In a preferred embodiment, step S2 includes: Design an isovariant reaction map encoder, (1) build a reaction characterization learning mechanism based on atomic mapping, and (2) use a geometric graph neural network structure to encode three-dimensional conformation data, wherein the GNN structure captures geometric information based on the UniMol Layer skeleton.

[0070] The reaction characterization learning mechanism in the isomorphic graph encoding module is characterized by the use of atom-mapping to explicitly model the alignment relationship between reactants and products. For both reactants and products, two different MPNN networks are used for aggregation and updating. , Corresponding message passing mechanism, Indicates the aggregation method of node messages. This indicates the update method for any node. Finally, through atomic mapping relationships, after characterizing the corresponding atoms in the reactants / products, collaborative learning is performed to complete an additional interaction on the reactants / products to capture reaction-level information. = .

[0071] The traditional message-passing-based GNN (MPNN) encoding layer is replaced with a UniMol Layer. Its key feature is the introduction of geometric biases into the Transformer's self-attention mechanism to encode 3D information. This mechanism transforms the geometric relationships between atoms into correction terms for the attention score.

[0072] The calculation method for the geometric deviation self-attention mechanism is as follows: for any two atoms in the diagram... and The attention calculation during message passing is as follows: ,in, From atoms Passed to the atom The message; and They are atoms and The query and key vector; Self-attention calculation representing the standard; It is a geometric deviation term that explicitly encodes the atoms. and The three-dimensional geometric relationships between atoms are derived from the Euclidean distance between atoms and other geometric quantities. Specifically, to supplement sufficient geometric information to compensate for the low-precision three-dimensional coordinate calculations, the following geometric quantities are additionally calculated as follows: Input , It is a distance encoding function. It's a key corner. It is the torsion angle.

[0073] The construction of the sequence model word segmenter structure includes: segmenting molecular representations in the SMILES language system based on regular expressions, and adding delimiters and other symbols to the regular expression system to support the segmentation of reaction expressions. Chemical reaction expressions are segmented from SMILES expressions into a series of text sequences based on chemical language morphemes. The vocabulary provided by the Chemformer model is expanded by adding reaction delimiters ('>>') and other special symbols, while simultaneously increasing its embedding matrix. For chemical reaction expressions... The sequence is obtained by passing it through an expanded word segmenter and an embedding layer. The atomic sequences from the graph encoder are flattened according to the order in which the SMILES expression appears, resulting in the graph sequence. .

[0074] The graph embedding-text alignment mask mechanism involves designing a cross-attention mechanism mask for aligned tokens to integrate information from the graph embedding space into a text sequence. For the input chemical expression, it first labels different molecules. Simultaneously, for graph sequence nodes and text sequence tokens, molecular tags are assigned according to their respective molecules. or Construct a special mask. Its characteristic is that, for position , If and equal to ,or This is a separator.

[0075] Using cross-attention and residual connections for modality alignment: The implementation of the cross-attention mechanism is as follows: ,in .

[0076] The residual connection mechanism consists of the following: Where O represents the calculation result of the cross-attention mechanism.

[0077] Enhanced response The input will be fed into a pre-trained sequence encoder skeleton. This sequence encoder employs a Chemformer model based on the BART (Bidirectional Autoregressive Transformer) architecture. The augmented sequence... By performing feature interaction and deep extraction through multiple Transformer Encoder layers, the final response latent vector representation containing image-text dual-modality information is obtained. This provides high-dimensional feature support for subsequent energy barrier prediction and structure generation.

[0078] The structure used in this invention differs from the standard Chemformer structure. It can be used directly as input to the encoder without going through the standard Chemformer word embedding layer again.

[0079] The model pre-training process includes the following three pre-training objectives: A masked language model strategy is employed to randomly mask some of the tokens of reactants or products, constructing a corrupted input sequence. The goal is to maximize the probability that the model reconstructs the original SMILES sequence, and the pre-training loss function is defined. To predict the cross-entropy loss between the token and the real token: Sequence Generation Pre-training – Forward Reaction Prediction: Pre-training the reactant SMILES sequences... As input to the encoder, the decoder is required to autoregressively generate the corresponding product SMILES sequence. .

[0080] Maximizing the conditional probability of the model generating the correct product sequence given the reactant sequence is equivalent to minimizing the following loss: ,in, Is the product sequence in Real Token at all times It is a token that was generated previously.

[0081] Sequence generation pre-training—chemical reaction conditions: This aims to enable the model to infer the external environmental factors (such as solvents, catalysts, etc.) required to achieve the transformation from the structural changes of the reaction itself, thereby incorporating kinetic information into the reaction characterization. The reaction sequence is then used for pre-training. As input to the encoder, the decoder is required to autoregressively generate SMILES sequences corresponding to the reaction conditions. .

[0082] That is, to maximize the model's ability to generate the correct reaction condition sequence under given reaction sequence conditions. The conditional probability is equivalent to minimizing the following loss: ,in, Is the reaction condition sequence in Real Token at all times It is a token that was generated previously.

[0083] In a preferred embodiment, step S3 includes: For unified modal features Extract the feature vector at a specific location or perform mean pooling on all token outputs to obtain a fixed-length reaction-level feature vector. , to feature vector The input is fed into a downstream regression prediction network. This regression network consists of a multilayer perceptron, containing linear layers and nonlinear activation functions (such as ReLU), and ultimately outputs a scalar value, namely the predicted response energy barrier. Construct the loss function for the regression task. Mean squared error (MSE) is typically used to measure the difference between the predicted energy barrier and the actual energy barrier. The differences between them .

[0084] In a preferred embodiment, step S4 includes: For the pre-training phase, token prediction accuracy and perplexity on the validation set are used as evaluation metrics to monitor the model's mastery of chemical language. In the energy barrier prediction fine-tuning task, the following statistical metrics are used to evaluate the model's predictive performance: Mean squared error (MSE) measures the squared average of the prediction error and reflects the overall goodness of fit of the model.

[0085] Mean Absolute Error (MAE) It measures the average absolute deviation between predicted and actual values ​​and is less sensitive to outliers.

[0086] Root mean square error (RMSE). It is used to measure the deviation between the predicted value and the actual value. The unit is consistent with the energy barrier unit (e.g., kcal / mol), which makes it easy to intuitively evaluate the prediction accuracy.

[0087] In a preferred embodiment, step S5 includes: Using the established bimodal representation ( This addresses the most challenging problem of transition state structure generation and employs an isovariant diffusion model to predict the atomic coordinates of the transition state. The diffusion model uses a Markov chain to gradually diffuse random noise. Transformed into a molecular structure with physical meaning. .

[0088] Predictive Model An E(3) equivariant graph neural network is used as the backbone. The E(3) equivariance ensures that the physical quantities of atomic coordinates remain unchanged under rotation and translation operations in three-dimensional space, which is a prerequisite for generating a reliable molecular structure.

[0089] By minimizing prediction noise and actual noise The differences between them are used to train the structure generation module: .

[0090] Noise Prediction Skeleton The characteristic is that for any rotation Peaceful relocation The operation must meet the following requirements: This means input coordinates The rotation and translation (on the left) will cause noise / gradient in the output prediction. Rotate by an equal amount (right side).

[0091] Representing a unified response Injected as conditional information into the denoising process In this process, the structure generation is guided. This ensures that the generated TS structure is not only geometrically reliable, but also optimized based on predicted energy barrier kinetic information.

[0092] The key denoising steps of its diffusion model are: .

[0093] in, It is the first Noise data of the step, It is a parameterized denoising function. This is the dual-mode conditional characteristic after fusion. This conditional generation process can predict the relative coordinates of each atom in the transition state stage, starting from the reactant / product coordinates.

[0094] In a preferred embodiment, step S6 includes: It is used to evaluate the geometric accuracy and chemical plausibility of the three-dimensional structure of the transition state (TS) generated by the isovariant diffusion model, and consists of two parts: Root Mean Square Shift (RMSD) Evaluation: Calculating the predicted transition state atom coordinates Compared with the true transition state coordinates obtained using high-precision quantum chemical calculations The root mean square displacement between them.

[0095] To eliminate the effects of rigid rotation and translation, the Kabsch algorithm must be used to achieve optimal alignment of the two sets of coordinates before calculation. The RMSD calculation formula is as follows: ,in, The number of atoms involved in the reaction. Let be a rotation matrix. This is the translation vector. A low RMSD value (typically less than 0.5 Å) indicates a high degree of agreement between the predicted and actual structures.

[0096] Structural validity check: In addition to geometric coordinate errors, validity verification needs to be conducted in conjunction with prior chemical knowledge, including: Bond length and bond angle check: Verify that the length and angle of the key chemical bonds in the generated TS structure are within a reasonable physical range, and exclude non-physical atomic overlap or breakage.

[0097] Imaginary frequency verification: Verify whether the predicted structure contains only one imaginary frequency after Hessian matrix analysis. This is a necessary and sufficient condition for determining that the structure is a first-order saddle point (i.e., a transition state).

[0098] The above specific embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; the above specific embodiments can be modified by those skilled in the art without departing from the principles and spirit of the present invention, or some or all of the technical features can be equivalently replaced; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model, characterized in that, Includes the following steps: S1. Data Acquisition and Standardization: Acquire raw data of chemical reactions, including reaction expressions and three-dimensional conformation information, and form standardized input containing sequence information, geometric information and atomic mapping relationships through automated preprocessing; S2. Bimodal Feature Encoding and Fusion: Construct a pre-trained multimodal fusion model, which includes an isomorphic graph encoder and a pre-trained sequence encoder, to extract three-dimensional features and global chemical semantic features from the standardized input, respectively, and to align and fuse the two features through a cross-modal fusion mechanism to generate a unified chemical reaction representation vector; S3. Transition State Energy Barrier Prediction and Structure Generation: Based on the unified chemical reaction representation vector, the following two tasks are performed simultaneously: I. By using a regression prediction network, the predicted activation energy barrier of a chemical reaction is output; II. A three-dimensional atomic coordinate structure of the chemical reaction transition state is generated by using an isovariant diffusion model conditioned on the aforementioned unified representation.

2. The method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model according to claim 1, characterized in that, The implementation of the isomorphic map encoder in step S2 includes: Construct a reaction map based on atom mapping and use this mapping relationship to explicitly model the atomic correspondence paths between reactants and products; A geometry-aware graph neural network layer is employed, which explicitly introduces a geometric deviation term calculated from the three-dimensional coordinates between atoms in the message passing or self-attention mechanism. The geometric deviation term is a function of at least one of the geometric quantities, namely, the Euclidean distance between atoms, the bond angle, and the dihedral angle.

3. The method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model according to claim 2, characterized in that, The geometric deviation term The calculation method is as follows: in, It is an atom The Euclidean distance between j and j It is an atom j and The bond angle formed, It is an atom j The angle of twist formed by l, It is a distance encoding function. It is a key-angle encoding function. It is the torsion angle encoding function.

4. The method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model according to claim 1, characterized in that, The implementation of the cross-modal fusion mechanism in step S2 includes: Construct a graph-text alignment mask matrix that constrains the range of attention interactions between graph node features and text token features based on the molecular identity to which the atom belongs; Cross-attention computation is employed, using text features as queries and graph features as keys and values. The graph-text alignment mask matrix is ​​applied during the attention weight calculation process to achieve targeted injection of three-dimensional geometric information into the semantic sequence.

5. The method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model according to claim 1, characterized in that, Prior to step S3, a multi-task pre-training phase is included, which comprises at least two of the following three self-supervised tasks: (a) Masked language modeling: Randomly mask the reaction expression lexical units and train the model to reconstruct the original sequence; (b) Forward reaction prediction: Using the reactant sequence as input, the model is trained to generate product sequences through autoregression; (c) Response condition reasoning: Using the complete response sequence as input, the model is trained to generate a sequence describing the response conditions through autoregression.

6. The method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model according to claim 1, characterized in that, The automated preprocessing in step S1 includes: Atomic-level mapping relationships between reactants and products are solved to establish the corresponding paths of atoms in the reaction process; Based on the aforementioned atomic mapping relationship, the atomic order of the three-dimensional coordinates of the transition state is rearranged to match the atomic order of the standard molecular expression.

7. The method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model according to claim 1, characterized in that, The noise prediction network of the equivariant diffusion model in step S3 satisfies the E(3) equivariance, that is, for any rotation of the input coordinates... With translation Its output noise prediction satisfies: in, For noise prediction networks, for Noisy coordinates at any given moment Let be the unified chemical reaction representation vector.

8. The method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model according to claim 7, characterized in that, The denoising process of the isotropic diffusion model is defined by the following formula: in, , , For diffusion process parameters, It is standard Gaussian noise.

9. The method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model according to any one of claims 1-8, characterized in that, It also includes automated evaluation steps: For the predicted activation barrier, calculate at least one of the following indices: mean square error, mean absolute error, and root mean square error. For the generated three-dimensional atomic coordinates of the transition state, the root mean square displacement between them and the reference coordinates is calculated, and the chemical rationality is verified.

10. The method for jointly predicting chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model according to claim 9, characterized in that, The verification of chemical rationality includes: Check whether the bond lengths and bond angles of the chemical bonds in the generated structure are within the preset reasonable physical range; And / or, perform frequency analysis on the generated structure to verify whether it has only one imaginary frequency.

11. A joint prediction system for chemical reaction energy barriers and transition states based on a dual-modal reaction characterization architecture and diffusion model, characterized in that, include: The data preprocessing module is used to perform data standardization and input construction; The bimodal fusion representation module is used to perform bimodal fusion representation learning and output a unified chemical reaction representation vector; The joint prediction execution module is used to receive the unified chemical reaction representation vector and simultaneously execute the energy barrier prediction and structure generation steps.

Images