Catalyst prediction method and system based on molecular characterization contrast learning

Abstract

The application belongs to the technical field of biological information. A catalyst prediction method and system based on molecular representation contrast learning are provided. Catalyst features are extracted according to a catalyst molecular graph. Reactant features and product features are extracted according to a reactant molecular graph and a product molecular graph. The reactant features and the product features are fused to obtain catalytic reaction features. The catalyst features and the catalytic reaction features are normalized respectively. The dot product of the normalized catalyst features and the normalized catalytic reaction features is calculated and scaled by a temperature parameter to obtain a similarity score of the catalyst features and the catalytic reaction features. According to the similarity scores of different catalysts, a preset number of catalyst types with the maximum similarity score are obtained. The application significantly improves the accuracy and interpretability of catalyst prediction through deep molecular representation and contrast learning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics, and particularly relates to a catalyst prediction method and system based on molecular representation contrast learning. BACKGROUND

[0002] The statements in this section merely provide background technology related to the present application and do not necessarily constitute prior art.

[0003] Catalyst research and development is the core technology of chemical catalysis field to design efficient catalysts. It uses the method of combining data-driven and theoretical calculation, starting from the target reaction, through active site analysis, electronic structure regulation and other strategies, to gradually screen and optimize potential catalytic materials, and realizes the intelligent deduction from "reaction demand to catalyst design". In traditional research, the process of catalyst research and development is highly dependent on the "trial and error method". The optimization of noble metal catalysts requires hundreds of experimental verifications, which takes several months to several years, and is also limited by the experience and contingency of researchers. This high-cost and low-efficiency research and development mode seriously restricts the innovation speed of catalysts, and the high research and development threshold becomes the main obstacle to technological breakthrough. In order to break through the experimental bottleneck of traditional empirical "trial and error method", the quantum chemistry calculation method represented by density functional theory provides a theoretical basis for rational design by simulating microscopic parameters such as catalyst surface adsorption energy. At the same time, the catalyst aided design method based on expert system is to construct the structure-activity relationship model by means of existing knowledge base and reasoning mechanism. However, the simulation accuracy of density functional theory calculation on complex heterogeneous catalytic system is limited by the computing resources, and the time consumption of a single calculation can reach several weeks. While the expert system can quickly screen candidate materials and predict performance trends, but it is limited by the finiteness of the knowledge base and the insufficient generalization ability of the model based on the rule base, and it is difficult to cover a wide chemical space.

[0004] In recent years, the rapid development of artificial intelligence technology has brought revolutionary opportunities for catalyst prediction. As an important branch of machine learning, deep learning has shown strong data-driven modeling capabilities in the field of catalyst prediction. Deep learning technology does not rely on chemical reagents and experimental instruments in training and inference, but only uses existing experimental data in the literature. Compared with quantum chemical calculations, it significantly reduces the consumption of computing resources, saving the cost of manpower and resources in the catalyst prediction process. In addition, deep learning models do not completely rely on or even rely on the experience and knowledge of chemists, which can break through the limitations of chemists' experience and intuition, discover catalyst molecules with excellent physical and chemical properties and catalytic activity, and speed up the process of finding suitable catalysts. Among them, graph neural networks are particularly outstanding in processing catalyst molecular structure data. This method can directly act on the molecular graph structure, capture the topological relationship between atoms, and improve the generalization ability of the model. Through an end-to-end training framework, deep learning can not only reduce the dependence on manual feature selection, but also achieve efficient learning on large-scale data sets, providing more accurate prediction means for high-throughput screening and performance optimization of catalysts.

[0005] However, the current development of computer-aided catalyst design methods still faces multiple challenges: first, existing research mainly relies on commercial databases and limited experimental data to construct datasets, which limits the generalization ability of the model, and the availability and standards of data are not uniform across different studies, hindering the horizontal comparison and standardized promotion of the method; second, the representation of the existing method for the catalyst molecule mostly stays at the shallow encoding (such as one-hot encoding), which fails to fully exploit its structural characteristics and complex interaction with reactants / products, restricting the reaction mechanism analysis and performance prediction accuracy; finally, the view information of the catalyst reaction molecule is diverse, such as the Simplified Molecular-Input Line-Entry System (SMILES) and the molecular graph, each of which carries unique structural and semantic information. However, the existing method fails to fully integrate and exploit the relevance between different views, limiting the comprehensive capture and utilization of molecular features. SUMMARY

[0006] In order to solve the problems of the prior art, the present application provides a catalyst prediction method and system based on molecular representation contrast learning, which can significantly improve the accuracy and interpretability of catalyst prediction through deep molecular representation and contrast learning.

[0007] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0008] In a first aspect, the present application provides a catalyst prediction method based on molecular representation contrast learning.

[0009] A catalyst prediction method based on molecular representation contrast learning includes the following processes:

[0010] obtaining a reactant molecular graph, a product molecular graph and a catalyst molecular graph;

[0011] extracting catalyst features according to the catalyst molecular graph, extracting reactant features and product features according to the reactant molecular graph and the product molecular graph, and fusing the reactant features and the product features to obtain catalyst reaction features;

[0012] performing normalization processing on the catalyst features and the catalyst reaction features respectively, calculating the dot product of the normalized catalyst features and the normalized catalyst reaction features, and scaling the dot product through a temperature parameter to obtain a similarity score of the catalyst features and the catalyst reaction features;

[0013] obtaining a preset number of catalyst types with the largest similarity scores according to the similarity scores of different catalysts.

[0014] In an implementation form of the first aspect of the present application, the reactant molecular graph, the product molecular graph and the catalyst molecular graph are all represented as a four-tuple , is an atomic node set, is a directed edge set, is a node feature matrix, is an edge attribute matrix.

[0015] In an implementation form of the first aspect of the present application, the catalyst features are extracted according to the catalyst molecular graph, including:

[0016] extracting initial representations of each atom in the catalyst molecular graph;

[0017] adopting a graph neural network model to transfer information between nodes and aggregate neighbor features according to the initial representations to obtain updated feature representations of each atom;

[0018] adopting a global pooling manner to aggregate the feature representations of all atoms in the same molecule to obtain a molecular feature representation;

[0019] sequentially passing the molecular feature representation through a full connection layer, a layer normalization processing and a Dropout processing to obtain the catalyst features.

[0020] In an implementation form of the first aspect of the present application, the reactant features and the product features are fused by adopting a recurrent neural network model to obtain the catalyst reaction features.

[0021] In an implementation form of the first aspect of the present application, the optimization target is a weighted sum of a contrastive loss and a cross-entropy loss, and the contrastive loss is: wherein, represents the number of catalyst samples; represents a loss scaling factor; represents the average log probability; represents the i-th catalyst sample; represents the i-th catalyst sample; represents the log probability.

[0022] As a further limitation of the first aspect of the application, the average log probability is calculated according to the positive sample pair mask and the similarity score, including: wherein, represents the catalyst sample and the catalyst sample represents the positive sample pair mask of the catalyst sample , represents the similarity score of the catalyst sample and the catalyst sample represents the similarity score of the catalyst sample is a minimum value, represents the similarity score of the catalyst sample and the catalyst sample represents the similarity score of the catalyst sample represents the log probability of the catalyst sample and the catalyst sample represents the log probability of the catalyst sample

[0023] As a further limitation of the first aspect of the application, the cross-entropy loss is: wherein, represents the i-th catalyst sample; represents the i-th catalyst sample; represents the feature similarity of the i-th catalyst sample in the correct class represents the similarity score of the catalyst sample and the catalyst sample represents the similarity score of the catalyst sample represents the exponential function.

[0024] In a second aspect, the application provides a catalyst prediction system based on molecular representation contrast learning.

[0025] A catalyst prediction system based on molecular representation contrast learning, comprising:

[0026] A molecular graph acquisition unit configured to acquire a reactant molecular graph, a product molecular graph, and a catalyst molecular graph;

[0027] A feature extraction unit configured to perform feature extraction according to the catalyst molecular graph to obtain catalyst features, extract reactant features and product features according to the reactant molecular graph and the product molecular graph, and fuse the reactant features and the product features to obtain catalyst reaction features;

[0028] ​The similarity score generation unit is configured to normalize the catalyst features and the catalytic reaction features respectively, calculate the dot product of the normalized catalyst features and the normalized catalytic reaction features, and scale the dot product by a temperature parameter to obtain a similarity score of the catalyst features and the catalytic reaction features.

[0029] The catalyst type prediction unit is configured to obtain a preset number of catalyst types with the maximum similarity scores according to the similarity scores of different catalysts.

[0030] In a third aspect, the present application provides a computer device, comprising: a processor and a computer readable storage medium.

[0031] The processor is adapted to execute the computer program.

[0032] The computer readable storage medium stores the computer program, and the computer program is executed by the processor to realize the catalyst prediction method based on molecular representation contrast learning according to the first aspect of the present application.

[0033] In a fourth aspect, the present application provides a computer readable storage medium storing a computer program, and the computer program is adapted to be loaded and executed by the processor to realize the catalyst prediction method based on molecular representation contrast learning according to the first aspect of the present application.

[0034] Compared with the prior art, the present application has the following beneficial effects:

[0035] The present application realizes a double breakthrough of catalyst prediction accuracy and model interpretability through deep integration of deep molecular representation technology and contrast learning framework. Traditional catalyst prediction methods mostly rely on manual features or shallow models, and it is difficult to capture complex chemical information of molecular structure. However, the present application uses graph neural network (GNN) to extract deep features of molecular graph, which can automatically learn high-dimensional representation of atomic arrangement, bonding mode and electronic structure in molecules. For example, through multi-layer graph convolution operation, the model can capture key structural features such as the conjugation effect of the pi bond in the benzene ring or the coordination unsaturation on the surface of the metal catalyst. These features are often lost due to dimension compression in traditional methods.

[0036] Through the contrast learning mechanism, the present application constructs positive and negative sample pairs of reactant-catalyst feature space, forces the model to distinguish the potential correlation between effective catalysts and ineffective catalysts, and through feature visualization, the matching mode of catalyst active sites and reactant functional groups can be clearly observed. This optimization strategy based on contrast loss not only improves the prediction accuracy, but also provides an interpretable chemical basis for catalyst design through feature similarity analysis.

[0037] The application fully retains the complete information of molecular atomic composition and chemical bond structure through a hierarchical feature extraction strategy; traditional molecular encoding methods (such as Morgan fingerprint) usually simplify the molecular structure into a fixed length vector, resulting in the loss of key information such as spatial configuration and electronic effect, the application adopts a graph attention network (GAT) to dynamically aggregate the features of the molecular graph, and the feature vector of each atom contains its chemical type, degree, formal charge and interaction information of adjacent atoms; for example, when processing platinum-based catalysts, the model can accurately identify the coordination number of surface platinum atoms and the adsorption configuration with the reactant molecules, and this multi-scale feature expression is closer to the real catalytic process than traditional methods.

[0038] Advantages of the additional aspects of the application will be partially given in the following description, partially will become obvious from the following description, or will be understood by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0039] The drawings accompanying the specification of this application form a part thereof, serve to further understand the application, and together with the description, explain the illustrative embodiments of the application, and do not constitute an improper limitation of the application.

[0040] Figure 1 A schematic diagram of the framework of the catalyst prediction model provided for an exemplary embodiment of the application;

[0041] Figure 2 A performance column chart when input data screening and model component changes are performed for an exemplary embodiment of the application;

[0042] Figure 3 Performance comparison results for an exemplary embodiment of the application;

[0043] Figure 4 A feature dimension reduction t-SNE graph for 10 randomly selected catalytic reactions for an exemplary embodiment of the application;

[0044] Figure 5 A randomly selected catalytic reaction case for an exemplary embodiment of the application;

[0045] Figure 6 A schematic diagram of the catalyst prediction system based on molecular representation contrast learning for an exemplary embodiment of the application;

[0046] Figure 7 A schematic diagram of a computer device for an exemplary embodiment of the application. DETAILED DESCRIPTION

[0047] The application will be further described below in conjunction with the drawings and embodiments.

[0048] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0049] The present implementation proposes a catalyst prediction method based on molecular representation contrastive learning, as shown in Figure 1 As shown in the figure, a CataCon model framework (i.e., a combined model framework of Catalyst and Contrastive Learning) is proposed, first, a four-tuple form molecular graph of reactants, products and catalysts is obtained, then feature extraction is performed through a graph neural network, the features of reactants and products are fused through a recurrent neural network to obtain catalyst reaction features, the similarity scores are calculated after normalizing the two types of features, a preset number of catalyst types are selected according to the scores, and the weighted sum of the contrastive loss and the cross-entropy loss is taken as the optimization objective (combined with the similarity matrix and the mask matrix), finally, the prediction classification result is obtained, which solves the problems of poor generalization ability and shallow molecular representation of traditional methods, and improves the accuracy and interpretability of catalyst prediction.

[0050] More specifically, the following processes are included:

[0051] S101: Construct a molecular graph data structure.

[0052] In order to solve the non-uniqueness problem of SMILES representation, the present application normalizes the SMILES strings of the USPTO-Catalyst dataset. First, for a SMILES combination containing multiple molecules, it is split according to the intermolecular separator to decompose it into independent molecular representations; then, each molecular structure is standardized to eliminate the differences caused by different writing methods and reduce the bias caused by resonance structures or different representations of molecular groups; then, the structure-standardized molecules are converted into unique and standardized SMILES forms to ensure that the same molecule always corresponds to a unique SMILES expression in different environments; finally, in order to ensure the consistency and comparability of the data, the SMILES of multiple molecules in the same group are sorted and recombined, and after cleaning up the repeated reaction data, a standardized SMILES dataset is constructed, thereby reducing redundancy and improving the model's ability to learn molecular similarity.

[0053] To reduce the impact of category mislabeling on model learning, the present application screens and cleans the catalyst categories in the data set. First, the SMILES representation of common solvent molecules is collected and normalized to ensure the accuracy of subsequent screening and selection. Then, the SMILES of each group of catalysts in the catalyst category is split into independent molecular representation, and compared with the collected solvent molecule SMILES one by one, and the compound combination composed of only solvent molecules is screened out and removed from the catalyst category. Given that the chemical properties of different catalysts in different solvent systems may change, this process only ensures that the catalyst category is not completely composed of solvent molecules, and retains these compound combinations as catalyst categories to construct a more accurate catalyst category data set, thereby reducing the model bias caused by solvent mislabeling and improving the prediction accuracy of the model for catalyst categories. After the two steps of SMILES normalization and catalyst category screening, the data information change in the data set is shown in Table 1.

[0054] Table 1: Data volume before and after cleaning of the USPTO-Catalyst data set

[0055]

[0056] In chemical informatics, molecular graphs avoid the limitations of sequence data flattening structural information by preserving topological connectivity, and are widely used in molecular representation, property prediction, reaction analysis and other fields. In order to more effectively utilize molecular information for training and prediction of deep learning models, the SMILES string is converted into an attribute-containing molecular graph.

[0057] In the present application, the molecular graph is represented as a four-tuple , which is defined as follows:

[0058] (1) is the set of atomic nodes, is the total number of atoms, representing non-hydrogen atoms in the molecule, represents the th atom;

[0059] (2) is the set of directed edges, which describes the chemical bond connection between atoms, each edge corresponds to a chemical bond and its inverse edge each edge ensures the symmetry of the graph structure;

[0060] (3) is the node feature matrix, which encodes the atomic-level chemical properties, where the dimension represents nine features including atomic number, chirality, total degree, formal charge, number of hydrogen atoms, number of free radical electrons, hybrid state, aromaticity and ring assignment;

[0061] (4) This is the edge attribute matrix, which describes the properties of chemical bonds, with dimensions... This represents three bond properties: bond type, stereochemistry, and conjugation, for each chemical bond. , Represents the set of chemical bonds, with bidirectional edge properties. Through mapping function calculate:

[0062] (1);

[0063] in, For key The original properties, For the first A dictionary of discrete values ​​corresponding to each chemical bond attribute. It represents the set of integers.

[0064] In step S101, the SMILES string is converted into a standardized molecular graph representation. The resulting molecular graph can be input into a graph neural network for molecular feature extraction, providing high-fidelity molecular structure data representation for subsequent tasks.

[0065] S102: Extracting catalytic reaction characteristics.

[0066] This module first extracts features from the molecular maps of reactants, products, and catalysts, and then merges the features of reactants and products to obtain the catalytic reaction features.

[0067] The first step is molecular graph feature extraction, based on the aforementioned molecular graph data representation. The molecular feature extraction module mainly includes the following parts: atomic feature representation learning, bond feature representation learning, graph structure feature extraction, global pooling, and fully connected layers.

[0068] The first part is atomic feature representation learning. Each column of the atomic feature matrix (corresponding to a discrete attribute) is converted into a vector through the corresponding embedding layer. Then, the embedding vectors are summed to form the initial representation of the atom. For each atom... , its first Attributes It will be through the corresponding embedding matrix Convert to vector representation, where This is the number of possible values ​​for this attribute. Represents the dimension. The atom is obtained by summing the embedding vectors of all attributes. initial representation :

[0069] (2);

[0070] where, represents the number of attributes.

[0071] The second part is the bond feature representation learning, which maps the discrete attributes of each chemical bond to a continuous vector. For each edge corresponding chemical bond , its attribute vector is given by the original attributes , which are obtained by a pre-defined mapping to get the normalized representation.

[0072] Each attribute is added after the corresponding embedding layer to form the vector representation of the edge:

[0073] (3);

[0074] where, represents the embedding matrix, represents the value of the th original discrete attribute of the edge .

[0075] The third part is the graph structure feature extraction, which uses the Graph SAGE graph neural network model to pass information between nodes and aggregate neighbor features through the graph structure E, thereby updating the representation of each atom (node) layer by layer.

[0076] For the th layer, the update formula of the representation of node is:

[0077] (4);

[0078] where, is the representation of the node in the previous layer; represents the neighbor set of node ; is the learnable weight matrix of the th layer; is a nonlinear activation function, represents the representation of the node in the th layer, represents the sequence number of the th node, represents the representation of the node in the previous layer.

[0079] Since the output of the graph neural network is the feature representation of each atomic node, in order to obtain the feature vector at the molecular level, the representation of all atoms within the same molecule needs to be aggregated.

[0080] The global mean pooling method is used:

[0081] (5);

[0082] wherein, represents a set of atoms within a molecule; is the updated node representation of the layer graph neural network, represents the global mean pooling result.

[0083] In order to further normalize and transform the molecular representation after pooling, finally a fully connected layer is processed, and the calculation process is as follows:

[0084] (6);

[0085] wherein, represents the result processed by a fully connected layer, represents the weight matrix of the fully connected layer; represents the bias of the fully connected layer.

[0086] Then, layer normalization (LN) and operations are performed:

[0087] (7);

[0088] (8);

[0089] The final output of the is the embedding representation of each molecule (i.e., the subsequent catalyst feature ), represents a set of real numbers, represents the dimension.

[0090] Then, the features of the reactants and products are fused to construct the catalytic reaction features (i.e., the subsequent catalyst feature ), the features of the reactants and products are added with sequence dimensions, and are spliced to form a sequence The obtained sequence is input into a GRU (Gated Recurrent Unit) model for fusion. GRU is a recurrent neural network unit with a gating mechanism, which is used to capture the timing information in the sequence. Here, GRU is used to fuse the feature information of the reactants and products in two steps.

[0091] For the input sequence , the GRU is updated at each time step according to the following formula:

[0092] (9);​​

[0093] (10);

[0094] (11);

[0095] (12);

[0096] where, is a Sigmoid function, denotes element-wise multiplication, represents the characteristics of the reactants, represents the characteristics of the products, represents the output of the update gate, represents the Sigmoid activation function, and represent the weight matrix of the reset gate, represents the bias term of the reset gate, represents the output of the reset gate, represents the candidate hidden state, and represent the weight matrix of the candidate hidden state, represents the bias term of the candidate hidden state, represents the hyperbolic tangent activation function, represents the hidden state of the current time step , and represents the hidden state of the previous time step .

[0097] The GRU processes the input sequence and outputs the hidden state of the entire sequence. The hidden state of the last time step is taken as the final fused feature, which comprehensively represents the information of the reactants and products, providing accurate reaction feature representation for downstream tasks. This method not only retains the advantage of graph neural networks in capturing molecular graph structure information, but also utilizes the time series modeling capability of GRU to effectively fuse the features of reactants and products.

[0098] S103: Capture the potential association between catalytic reactions and catalysts.

[0099] The structural information of the catalyst is integrated into the overall feature learning process instead of the traditional one-hot encoding method, thereby obtaining a more discriminative representation. A loss function is designed, and its basic process is as follows. First, normalize the catalytic reaction features and catalyst features obtained by the molecular feature extraction module, denoted as:

[0100] ​ (13);

[0101] (14);

[0102] in, , This indicates the number of catalyst samples.

[0103] Subsequently, the cosine similarity between the two was calculated, and the similarity scores of the catalytic reaction characteristics and catalyst characteristics were obtained by scaling the results using the dot product method and the temperature parameter τ, thus yielding a similarity matrix. :

[0104] (15);

[0105] in, Representing the The catalyst sample and the first Similarity of individual catalyst samples.

[0106] Next, positive and negative samples are constructed for contrastive learning. Samples of the same catalyst category are considered positive samples, and samples of different catalyst categories are considered negative samples. A mask matrix for positive sample pairs is then constructed. Its elements are defined as: when the sample With sample When they belong to the same category ,otherwise For each pair of samples Calculate its logarithmic probability :

[0107] (16);

[0108] in, It is a very small quantity used to prevent numerical instability. Representing the The serial number of each sample. Representative sample With sample The similarity.

[0109] For each sample The average log probability of positive sample pairs for:

[0110] (17);

[0111] Ultimately, the contrastive learning loss for the entire batch is defined as:

[0112] (18);

[0113] in, represents the number of catalyst samples; is the loss scaling factor. By introducing the structural features of catalysts and combining with the mask matrix and the design of the similarity matrix , the model pulls the representations of similar samples closer in the high-dimensional feature space while pushing the representations of dissimilar samples further apart, thereby enhancing the model's ability to capture the intrinsic correlation between catalytic reactions and catalysts. Compared with the traditional method of using only One-Hot Encoding, this module makes the information expression more rich by introducing the structural features of catalysts, improving the robustness and generalization performance of prediction.

[0114] S104: Predict the catalyst category.

[0115] The purpose of the prediction and classification module is to accurately predict the catalyst category using the feature representation learned from the contrastive learning module. In this study, the Cross-Entropy Loss function is used to optimize the classification task. First, as in the contrastive learning module, the catalyst reaction features and catalyst structure features obtained from the molecular graph feature representation module are normalized to eliminate scale differences, and the feature similarity logits are calculated. To ensure numerical stability, the logits are normalized by the maximum value, and the prediction probabilities of each category are obtained by SoftMax operation :

[0116] (19);

[0117] wherein represents the exponential function, represents the similarity matrix of the th catalyst sample and the th catalyst sample, represents the maximum value of similarity.

[0118] In the actual prediction process, the strategy is often used to measure the recognition effect of the model. For example, for a given value, the model will select the top category with the highest prediction probability, and then compare it with the actual label to calculate the accuracy.

[0119] To optimize the classification task, the multi-class Cross-Entropy Loss function is used, and the mathematical expression of the loss function is:

[0120] (20); ​

[0121] wherein, represents the number of catalyst samples, represents the logits value of the i-th catalyst sample on the correct category, The loss of the current sample is small if the probability of the target category is high, i.e. the prediction is accurate, and vice versa, if the probability of the target category is low, i.e. the prediction is inaccurate.

[0122] In this model framework, the overall optimization objective function is jointly determined by the contrastive loss and the cross-entropy loss , i.e.:

[0123] (21).

[0124] wherein, represents the weight coefficient.

[0125] The two are summed by weighting to form the total loss function, wherein the weight coefficient controls the relative contribution of the two parts of the loss. This coefficient is a hyperparameter that can be adjusted during training to achieve optimal performance configuration. Such a design enables the model not only to learn the discriminative representation between catalyst reaction data samples, i.e. to pull similar samples and push dissimilar samples, but also to improve the accuracy of category prediction, thereby enhancing the generalization ability of the model.

[0126] S105: Verify the performance of the CataCon model.

[0127] In order to evaluate the prediction ability of the CataCon model, the present application sets up a contrast experiment for feature extraction and a contrast experiment for contrast learning according to the division of the above steps.

[0128] As shown in Figure 2 , the following ten control groups are used in the contrast experiment: the present application scheme, the present application scheme + improved graph attention network (i.e. using an improved graph attention network to replace the feature extraction network in the present application scheme), the present application scheme + graph attention network (i.e. using a graph attention network to replace the feature extraction network in the present application scheme), the present application scheme + graph reconstruction network (i.e. using a graph reconstruction network to replace the feature extraction network in the present application scheme), the present application scheme + graph convolution network (i.e. using a graph convolution network to replace the feature extraction network in the present application scheme), negative sampling correction contrast learning + graph sampling aggregation; negative sampling correction contrast learning + improved graph attention network, negative sampling correction contrast learning + graph attention network, negative sampling correction contrast learning + graph reconstruction network, negative sampling correction contrast learning + graph convolution network.

[0129] ​​More specifically, the feature extraction part adopts four other graph neural network architectures, namely graph convolution network (GCN), graph isomorphism network (GIN), graph attention network (GAT_v1) and improved graph attention network (GAT_v2). The contrast learning part compares the loss function proposed in the application with the negative sampling correction contrast learning (PUCL). The experiment adopts an 8:1:1 dataset segmentation ratio to divide the original dataset into three independent sets of training set, validation set and test set, so as to eliminate potential bias in the evaluation process. Such a division strategy can fully utilize the overall distribution characteristics of the data and effectively reduce the risk of overfitting, so as to obtain more reliable experimental results in the test stage. Subsequently, the performance of each model under different conditions will be analyzed in detail, and the potential factors affecting the performance will be discussed. The core goal of the application is to identify the model with the best performance in a specific task through direct experimental comparison, so as to provide valuable reference and inspiration for future deep learning research.

[0130] The experimental results are shown in Table 2 and Figure 2 In the comparative experiment of the feature selection module, CataCon of the application adopts Graph SAGE as the feature extraction method, and in the Top-1 (prediction accuracy of the first ranking), Top-3 (prediction hit rate of the top three), Top-5 (prediction hit rate of the top five) and Top-10 (prediction hit rate of the top ten) four evaluation indexes, the best performance is obtained, which is 1.25%, 0.08%, 0.31% and 0.13% higher than the second ranked model combination respectively. In addition, in the comparative experiment of the contrast learning module, CataCon compared with the PUCL method, the corresponding indexes are improved by 13.51%, 9.78%, 7.2% and 6.49% respectively. The above results show that compared with the existing contrast learning method, the proposed contrast learning loss function can pay more attention to the distinguishing information between positive and negative samples, thereby improving the discrimination ability of the model. On the other hand, GraphSAGE can capture the topological relationship and feature information of atoms and their adjacent atoms in molecules through neighbor sampling and feature aggregation mechanism, and then represent the local chemical environment of atoms, which makes this method perform well in complex molecular graph modeling, and can efficiently learn the molecular structure information of reactants, products and catalyst combinations.

[0131] Table 2: CataCon model performance comparison results

[0132]

[0133] In summary, the comparative learning module designed in the application has better optimization effect than PUCL, and GraphSAGE performs better than the second-ranked model in feature extraction, especially in Top-1 accuracy. These results prove that the combination of feature extraction and comparative learning module proposed in this study exhibits better performance in the direct prediction task of catalyst categories.

[0134] The application designs ablation experiments from two angles of data and model architecture to analyze the role of key components in depth. From the data angle, the molecular information of reactants and products is shielded respectively, and the single reactant or product feature is used as the feature input of the catalytic reaction for subsequent training to evaluate the contribution of different molecular information to the model performance.

[0135] From the model architecture angle, two ablated model variants are designed for comparative analysis. One model variant removes the GRU model used to fuse the reactant and product features to construct the catalytic reaction feature in the molecular feature extraction module, and instead directly reduces the spliced reactant-product feature through a linear layer to generate the final catalytic reaction feature for subsequent training. The other model variant removes the contrast learning mechanism of the reaction feature and the catalyst feature in the contrast learning module, and instead uses a fully connected layer as a replacement for the module. This ablation analysis helps to quantify the influence of different modules on the overall performance of the model and provides guidance for optimizing the network structure. The results are referred to Table 3.

[0136] Table 3: Performance of CataCon model under different parts of input data shielding and model component change

[0137]

[0138] From Figure 3 It can be seen that the performance of the original complete model is better than the simplified version in all evaluation indicators, which verifies that each component is indispensable for achieving the best performance. The ablation experiment results clearly reveal the contribution of different input features and model components to the overall performance. From Figure 3As can be seen from the data, the unmodified model performs best on all Top-k accuracy indicators, indicating that each key component contributes to the final prediction performance. In the input data ablation experiment, using only reactant or product information separately results in a significant decline in model performance, especially after masking the reactants, the Top-1 accuracy drops to 0.0023. Similarly, after masking the product information, the model performance is slightly higher than that of masking the reactants, but it is still much lower than that of the complete model, indicating that both reactant and product information are necessary parts of constructing effective catalytic reaction features. In the model architecture ablation experiment, the Top-1 accuracy of the model after removing the GRU is only 0.0014, indicating that the GRU plays a key role in integrating reactant and product features, and its absence causes the model to fail to effectively construct catalytic reaction features. In addition, after removing the contrast learning mechanism, the Top-1 accuracy drops to 0.0036, and the Top-3 and above accuracy is relatively high, indicating that contrast learning plays an important role in optimizing the model's discriminative ability, especially in the low Top-k range. The experimental results in this section show that every component of the data and model architecture is crucial to the overall performance. Whether it is to remove part of the input features or adjust the network structure, any change will have a significant impact on the model performance.

[0139] To evaluate the feature representation ability of the CataCon model and improve its interpretability, t-distributed Stochastic Neighbor Embedding (t-SNE) is used to reduce the dimensionality of the output of the molecular feature extraction module of the CataCon model. The purpose is to explore the distribution of features learned by the model and analyze the similarity and difference between samples. Through this method, the clustering pattern of features of different categories in the embedding space can be revealed, thereby evaluating the discriminative ability and generalization performance of the model. In addition, t-SNE can also be used to detect whether the features extracted by the model have reasonable separability, providing an important reference for model optimization.

[0140] As Figure 4As shown, 10 samples of each category were randomly selected from the test set, with 10 data per category, and input into the trained model to obtain feature representations. Subsequently, the extracted high-dimensional feature matrix was reduced and visualized using t-SNE to more intuitively observe the distribution characteristics of the data and the potential patterns learned by the model. The t-SNE visualization results show the overall distribution of the catalyst prediction model based on contrastive learning after dimensionality reduction. The model forms a high-dimensional representation by concatenating the catalyst features with the reaction features (obtained by fusing the reactant features and product features), and uses t-SNE for dimensionality reduction to explore the inherent clustering structure. From the visualization results, samples of different categories form clear cluster-like distribution in the t-SNE space, indicating that the features learned by the model have good discrimination ability. The clustering tightness within the class reflects the consistency of the model on similar samples, while the significant separation between classes further proves the model's ability to identify different catalyst types. This clear clustering structure indicates that the contrastive learning strategy of the model plays a positive role in enhancing feature discrimination, allowing similar categories of catalytic reactions to be more closely aggregated in low-dimensional space, while samples of different categories maintain good separation. It is worth noting that the distribution of points in some categories is relatively tight, which may indicate that the consistency of this type of catalyst in the feature space is higher, while some categories show greater internal dispersion, which may be related to the diversity of catalysts in different reaction systems. This distribution feature not only provides a basis for further analyzing the clustering quality of specific categories, but also helps to evaluate the model's generalization ability on different types of catalysts.

[0141] Overall, the t-SNE results provide intuitive support for understanding the learning ability of the catalyst prediction model and provide important clues for model optimization and improvement. This visualization analysis not only helps to evaluate the actual performance of the model, but also further verifies the excellent performance of CataCon in feature extraction and sample classification, providing a powerful tool for future catalyst design and screening.

[0142] As Figure 5As shown, the present application randomly selects a catalytic reaction from the test set as a case for detailed analysis. The reaction takes 1-methyl-3-nitroindole ([CH3:1][C:2]1[C:3](N+:12](O-)=O)=[CH:4][CH:5]=[C:6]2[C:11]=1[N:10]=[CH:9][CH:8]=[CH:7]2) as the reactant, and together with formate (C([O-])=O) and ammonium ion ([NH4+]) to occur reduction amination reaction, generate the target product 1-methyl-3-aminoindole ([NH2:12][C:3]1[C:2]=[C:11]2[C:6]([CH:7]=[CH:8][CH:9]=[N:10]2)=[CH:5][CH:4]=1), in which the nitro group is reduced to amino group, and ammonium ion as a nitrogen source participates in the reaction. First, the SMILES strings of the reactant group, reagent and product group involved in the reaction are data cleaned and preprocessed, and are converted into molecular graph data of molecules, so as to more accurately capture the topological structure and chemical information inside the molecule. Subsequently, the processed molecular graph data is input into the prediction model, and the model outputs the candidate catalyst ranking of the corresponding reaction based on the feature extraction and deep learning method of the molecular graph.

[0143] Referring to Table 4, among the top ten candidate catalysts predicted by the model, the first-ranked is the ethanol-coordinated palladium catalyst "CCO.[Pd]", followed by the ethanol and carbon co-coordinated palladium catalyst "CCO.[C].[Pd]", the ethanol and water co-coordinated palladium catalyst "CCO.O.[Pd]", and so on. It is worth noting that the actual catalyst of this reaction is "CCO.[Pd]", which is consistent with the highest-ranked candidate of the model, indicating that the model can effectively identify the catalyst category suitable for the reaction. Further analysis of the candidate results can find that the model selects catalysts with certain similarity in ligand structure while keeping palladium ([Pd]) as the catalytic center. Candidates 2, 3, 4, 5, 6, and 8 all retain the ethanol group "CCO" or similar oxygen / carbon-containing groups, and adjust the electron density and spatial configuration by introducing carbon, oxygen, and other functional groups. For example, candidate 6 is a palladium complex containing double hydroxyl groups "CCO.[OH-].[OH-].[Pd+2]", and candidate 8 is an ethanol and ethyl acetate co-coordinated palladium structure "CCOC(C)=O.[Pd]". This trend shows that the model not only identifies the key structural features of the catalyst, but also provides modification schemes with potential catalytic activity while keeping the core structure unchanged. In addition, although there are certain differences in the ligand structures of the candidate catalysts, since they all have [Pd] as the catalytic center, their electronic transfer and coordination characteristics in the catalytic reaction may have high similarity. In particular, the catalytic activity and selectivity of the candidate catalysts with less change in ligand structure may be close to that of the actual catalyst, which provides a reference for further optimization of the catalytic system.

[0144] Table 4: Top ten candidate catalysts for the predicted catalytic reaction

[0145]

[0146] In summary, the catalyst prediction model not only accurately identifies the actual catalyst, but also selects candidate catalysts with potential catalytic activity based on structural similarity. This shows that the model has good applicability in catalyst design and screening, and can provide a reasonable candidate set for catalyst optimization and lay a foundation for subsequent experimental research.

[0147] Figure 6 A catalyst prediction system based on molecular representation contrast learning is shown, comprising:

[0148] The molecular graph acquisition unit 601 is configured to acquire a reactant molecular graph, a product molecular graph, and a catalyst molecular graph;

[0149] The feature extraction unit 602 is configured to perform feature extraction according to the catalyst molecule graph to obtain catalyst features, perform feature extraction according to the reactant molecule graph and the product molecule graph to obtain reactant features and product features, and fuse the reactant features and the product features to obtain catalyst reaction features.

[0150] The similarity score generation unit 603 is configured to perform normalization processing on the catalyst features and the catalyst reaction features respectively, calculate the dot product of the normalized catalyst features and the normalized catalyst reaction features, and scale the dot product by a temperature parameter to obtain a similarity score of the catalyst features and the catalyst reaction features.

[0151] The catalyst type prediction unit 604 is configured to obtain a preset number of catalyst types with the largest similarity scores according to the similarity scores of different catalysts.

[0152] It can be understood that the above-mentioned units can be combined into one or several other units respectively or all, or some of the units can be further split into a plurality of units with smaller functions to constitute, which can realize the same operation without affecting the realization of the technical effects of the embodiments of the present application. The above-mentioned units are divided based on logical functions, and in actual application, the functions of one unit can also be realized by multiple units, or the functions of multiple units can be realized by one unit. In other embodiments of the present application, the system can also include other units, and in actual application, these functions can also be realized by other units, and can be realized by multiple units in cooperation.

[0153] According to another embodiment of the present application, the system of the present embodiment can be constructed by running a computer program (including program code) capable of performing each step involved in the corresponding method of the present application on a general computing device such as a computer including processing elements and storage elements such as a Central Processing Unit (CPU), a Random Access Memory (RAM), a Read Only Memory (ROM), etc., the computer program can be recorded on a computer readable recording medium, and loaded into the above-mentioned computing device through the computer readable recording medium and run therein.

[0154] Figure 7 A computer device is shown, which includes a processor 701, a communication interface 702, and a computer readable storage medium 703. Wherein the processor 701, the communication interface 702 and the computer readable storage medium 703 can be connected through a bus or other means.

[0155] The communication interface 702 is configured to receive and send data, the computer readable storage medium 703 can be stored in the memory of the electronic device, the computer readable storage medium 703 is configured to store a computer program, the computer program includes program instructions, and the processor 701 is configured to execute the program instructions stored in the computer readable storage medium 703.

[0156] The processor 701 is a computing core and a control core of the electronic device, and is adapted to implement one or more instructions, and is specifically adapted to load and execute one or more instructions to implement a corresponding method flow or a corresponding function.

[0157] The processor 701 is configured to execute the following process:

[0158] Obtain a reactant molecular graph, a product molecular graph and a catalyst molecular graph;

[0159] According to the catalyst molecular graph, a catalyst feature is extracted, according to the reactant molecular graph and the product molecular graph, a reactant feature and a product feature are extracted, and the reactant feature and the product feature are fused to obtain a catalytic reaction feature;

[0160] The catalyst feature and the catalytic reaction feature are normalized respectively, the dot product of the normalized catalyst feature and the normalized catalytic reaction feature is calculated, and the dot product is scaled by a temperature parameter to obtain a similarity score of the catalyst feature and the catalytic reaction feature;

[0161] According to the similarity scores of different catalysts, a preset number of catalyst types with the maximum similarity score are obtained.

[0162] The application further provides a computer readable storage medium, which is a memory device in the electronic device and is used for storing programs and data. It can be understood that the computer readable storage medium herein can include a built-in storage medium in the electronic device, and of course can also include an expansion storage medium supported by the electronic device. The computer readable storage medium provides a storage space, and the storage space stores a processing system of the electronic device.

[0163] In addition, one or more instructions adapted to be loaded and executed by the processor are also stored in the storage space, and the instructions can be one or more computer programs (including program codes). It should be noted that the computer readable storage medium herein can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory; optionally, the computer readable storage medium can also be at least one computer readable storage medium located away from the aforementioned processor.

[0164] In one embodiment, the computer readable storage medium stores one or more instructions; the processor loads and executes the one or more instructions stored in the computer readable storage medium to implement the following process:

[0165] Obtaining a reactant molecular graph, a product molecular graph and a catalyst molecular graph;

[0166] According to the catalyst molecular graph, a catalyst feature is extracted; according to the reactant molecular graph and the product molecular graph, a reactant feature and a product feature are extracted; and the reactant feature and the product feature are fused to obtain a catalytic reaction feature;

[0167] The catalyst feature and the catalytic reaction feature are normalized respectively; the dot product of the normalized catalyst feature and the normalized catalytic reaction feature is calculated and scaled by a temperature parameter to obtain a similarity score of the catalyst feature and the catalytic reaction feature;

[0168] According to the similarity scores of different catalysts, a preset number of catalyst types with the largest similarity score are obtained.

[0169] Those of ordinary skill in the art can understand that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present application can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical solution. Those of ordinary skill in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0170] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in or transmitted by a computer readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired (for example, coaxial cable, optical fiber, digital line) or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer readable storage medium can be any available medium that can be accessed by a computer or a data processing device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (for example, floppy disk, hard disk, magnetic tape), optical media (for example, DVD), or semiconductor media (for example, solid state disk) and the like.

[0171] The above merely describes the preferred embodiments of the present application and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A catalyst prediction method based on molecular characterization comparative learning, characterized in that, Includes the following processes: Obtain molecular diagrams of reactants, products, and catalysts; Catalyst features are obtained by feature extraction based on catalyst molecular diagrams, reactant features and product features are obtained by extracting reactant features and product features based on reactant molecular diagrams and product molecular diagrams, and catalytic reaction features are obtained by fusing reactant features and product features. The catalyst characteristics and catalytic reaction characteristics are normalized respectively. The dot product of the normalized catalyst characteristics and catalytic reaction characteristics is calculated and scaled by the temperature parameter to obtain the similarity score between the catalyst characteristics and catalytic reaction characteristics. Based on the similarity scores of different catalysts, the catalyst type with the highest preset number of similarity scores is obtained; First, obtain the molecular diagrams of reactants, products, and catalysts in quaternary form, where the quaternary form... , For a set of atomic nodes, Let be a set of directed edges. The node feature matrix, The edge attribute matrix is ​​used; then, feature extraction is performed through graph neural network, and the catalytic reaction features are obtained by combining reactant and product features with recurrent neural network. After normalizing the two types of features, similarity scores are calculated, and a preset number of catalyst types are selected according to the scores. At the same time, the weighted sum of contrast loss and cross-entropy loss is used as the optimization objective, and finally the predicted classification results are obtained.

2. The catalyst prediction method based on molecular characterization comparative learning as described in claim 1, characterized in that, Catalyst features are obtained by feature extraction based on the catalyst molecular diagram, including: Extract the initial representation of each atom in the catalyst molecular diagram; Based on the initial representation, a graph neural network model is used to pass information between nodes and aggregate neighbor features to obtain the updated feature representation of each atom; Global pooling is used to aggregate the feature representations of all atoms within the same molecule to obtain the molecular feature representation; The molecular features were sequentially processed through a fully connected layer, layer normalization, and Dropout to obtain the catalyst features.

3. The catalyst prediction method based on molecular characterization comparative learning as described in claim 1, characterized in that, The catalytic reaction characteristics are obtained by fusing reactant and product characteristics using a recurrent neural network model.

4. The catalyst prediction method based on molecular characterization comparative learning as described in claim 1, characterized in that, The optimization objective is a weighted sum of the contrastive loss and the cross-entropy loss, where the contrastive loss is: ,in, This represents the number of catalyst samples; This represents the loss scaling factor; Represents the average logarithmic probability; Representing the One catalyst sample; This represents the logarithmic probability.

5. The catalyst prediction method based on molecular characterization comparative learning as described in claim 4, characterized in that, Mean log probability Calculated based on positive sample pairs, masks, and similarity scores, including: ,in, Representative catalyst sample With catalyst sample Positive sample pair mask, , Representative catalyst sample With catalyst sample Similarity score, It is a local minimum. Representative catalyst sample With catalyst sample Similarity score; Representative catalyst sample With catalyst sample The logarithmic probability.

6. The catalyst prediction method based on molecular characterization comparative learning as described in claim 4, characterized in that, The cross-entropy loss is: ,in, Representing the One catalyst sample; Indicates the first One in the correct category Feature similarity, Representative catalyst sample With catalyst sample Similarity score; This represents an exponential function.

7. A catalyst prediction system based on molecular characterization comparative learning, characterized in that, include: The molecular diagram acquisition unit is configured to acquire reactant molecular diagrams, product molecular diagrams, and catalyst molecular diagrams. The feature extraction unit is configured to: extract catalyst features based on the catalyst molecular map, extract reactant features and product features based on the reactant molecular map and product molecular map, and fuse the reactant features and product features to obtain catalytic reaction features; The similarity score generation unit is configured to: normalize the catalyst features and catalytic reaction features respectively, calculate the dot product of the normalized catalyst features and catalytic reaction features and scale it using the temperature parameter to obtain the similarity score between the catalyst features and catalytic reaction features; The catalyst type prediction unit is configured to: obtain the catalyst type with the highest preset number of similarity scores based on the similarity scores of different catalysts; First, obtain the molecular diagrams of reactants, products, and catalysts in quaternary form, where the quaternary form... , For a set of atomic nodes, Let be a set of directed edges. The node feature matrix, The edge attribute matrix is ​​used; then, feature extraction is performed through graph neural network, and the catalytic reaction features are obtained by combining reactant and product features with recurrent neural network. After normalizing the two types of features, similarity scores are calculated, and a preset number of catalyst types are selected according to the scores. At the same time, the weighted sum of contrast loss and cross-entropy loss is used as the optimization objective, and finally the predicted classification results are obtained.

8. A computer device, characterized in that, include: Processor and computer-readable storage media; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the catalyst prediction method based on molecular characterization comparative learning as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1 to 6, for the catalyst prediction method based on molecular characterization comparative learning.

Images