Multi-modal chemical reaction yield prediction method based on adaptive data screening

Abstract

The invention discloses a multi-mode chemical reaction yield prediction method based on adaptive data screening, and belongs to the technical field of chemical synthesis and machine learning crossing. According to the method, a multi-modal input system containing one-dimensional chemical attribute data, a two-dimensional molecular structure map and a three-dimensional spatial configuration is constructed, so that multi-dimensional analysis of molecular interaction is realized; and meanwhile, a multi-stage training process is adopted to simulate a human cognitive rule, so that the model progressively learns layer by layer from a basic reaction feature to a complex reaction mechanism. According to the method, the screening threshold is dynamically adjusted, noise interference is effectively suppressed, and the adaptability to long-tail distribution data is improved. Compared with a traditional single-mode prediction model, the method has the advantages that the prediction accuracy on a noisy data set is improved, the trial and error cost of chemical synthesis experiments is remarkably reduced, and reliable technical support is provided for efficiently screening reaction conditions and accelerating research and development of new compounds.

Description

Technical Field

[0001] The present invention belongs to the field of intersectional technology involving chemistry and artificial intelligence, and in particular relates to a multimodal chemical reaction yield prediction method based on adaptive data screening. Background Art

[0002] Chemical reaction yield prediction is a key technology in the chemical industry, crucial for optimizing production processes, reducing production costs, and improving product quality. Traditional methods typically rely on experimental data and empirical models, which are not only inefficient but also often struggle to accurately predict yields when faced with complex chemical reactions. With the advancement of cheminformatics and machine learning technologies, data-driven yield prediction models have gradually gained attention. However, existing yield prediction models still face numerous challenges in handling high-noise data and fusing multimodal information.

[0003] First, there are often a large amount of noise and outliers in chemical reaction data, which will seriously interfere with the training of the model, making it difficult for the model to converge or fall into a local optimal solution. Secondly, chemical reactions involve data of multiple modalities, such as one-dimensional SMILES sequences, two-dimensional molecular graphs, and three-dimensional molecular conformations. There are complex correlations between these modal information. How to effectively fuse this information to improve the predictive ability of the model is an urgent problem to be solved. In addition, when processing multimodal data, existing models often lack the alignment and fusion mechanism of different modal features, resulting in the model being unable to fully utilize the advantages of multimodal information. Therefore, there is an urgent need for a chemical reaction yield prediction model that can effectively resist data noise interference and deeply fuse multimodal information to improve the accuracy and reliability of the prediction. Summary of the Invention

[0004] This paper aims to provide a multimodal chemical reaction yield prediction method based on adaptive data screening to address the shortcomings of existing chemical reaction yield prediction models in handling high-noise data and integrating multimodal information. By introducing an adaptive data screening mechanism and a cross-modal comparative learning strategy, the present invention effectively resists the interference of data noise and deeply integrates multimodal information, thereby improving the accuracy and reliability of chemical reaction yield prediction.

[0005] In order to achieve the above-mentioned object of the invention, the present invention specifically adopts the following technical solutions:

[0006] In a first aspect, the present invention provides a multimodal chemical reaction yield prediction method based on adaptive data screening, comprising the following steps:

[0007] S1. Obtain a one-dimensional SMILES sequence and its corresponding two-dimensional molecular graph to form an unlabeled multimodal dataset;

[0008] S2. In the pre-training phase, the network parameters of the multi-layer perceptron predictor are fixed, and the micro-multimodal encoder is trained on the unlabeled multimodal dataset using a cross-modal contrastive learning training strategy.

[0009] S3. After pre-training is complete, the unlabeled multimodal dataset is labeled to obtain a labeled yield dataset. The network parameters of the multi-layer perceptron predictor are unfrozen, and a microscopic multimodal chemical reaction yield prediction model is constructed from the microscopic multimodal encoder and the multi-layer perceptron predictor. The microscopic multimodal chemical reaction yield prediction model is end-to-end jointly trained based on the labeled yield dataset. The weights of the microscopic multimodal chemical reaction yield prediction model are synchronously updated through backpropagation to obtain a trained microscopic multimodal chemical reaction yield prediction model.

[0010] During the joint training process, an adaptive data screening algorithm dynamically removes abnormal training data based on prediction confidence and feature space distribution, iteratively updates the labeled yield dataset, and calculates the variance estimate of the chemical reaction yield prediction value through the uncertainty quantification module. Confidence intervals for the chemical reaction yield prediction value are generated based on Monte Carlo sampling.

[0011] S4. The one-dimensional SMILES sequence to be used for yield prediction and its corresponding two-dimensional molecular graph are input into the trained microscopic multimodal chemical reaction yield prediction model. The microscopic multimodal encoder extracts high-dimensional fusion features, and the multi-layer perceptron predictor maps the extracted high-dimensional fusion features into chemical reaction yield prediction results. The uncertainty quantification module calculates the variance estimate of the chemical reaction yield prediction results, and the confidence interval of the chemical reaction yield prediction results is generated based on Monte Carlo sampling, thereby realizing multimodal chemical reaction yield prediction based on adaptive data screening.

[0012] Based on the above solution, each step can be implemented in the following preferred specific manner.

[0013] As a preferred embodiment of the above-mentioned first aspect, in the microscopic multimodal encoder of step S2, the one-dimensional SMILES sequence is encoded by the SMILES sequence encoder, and the one-dimensional SMILES sequence is mapped to the vector space to obtain a SMILES feature vector, the two-dimensional molecular graph is encoded by the molecular graph encoder to generate a global graph representation of the two-dimensional molecular graph, and a three-dimensional molecular conformation feature is generated based on the two-dimensional molecular graph. Then, the SMILES feature vector, the global graph representation of the two-dimensional molecular graph and the three-dimensional molecular conformation feature are combined through a feature splicing operation to generate a comprehensive feature vector.

[0014] As a preferred embodiment of the above-mentioned first aspect, the specific generation method of the three-dimensional molecular conformational characteristics is as follows: first, the three-dimensional molecular initial conformation is obtained from the two-dimensional molecular graph using a technology related to conformation prediction, and the energy of the three-dimensional molecular initial conformation is minimized by the MMFF molecular mechanics force field to obtain the corresponding stable conformation and form a series of three-dimensional molecular stable conformation sets participating in chemical reactions; then the chemical informatics and machine learning toolkit RDKit is used to calculate the stable conformation set to obtain the three-dimensional molecular conformation characteristics including geometric diameter, moment of inertia, and mass distribution, or the three-dimensional molecular conformation encoder is used to encode the stable conformation set to generate three-dimensional molecular conformation characteristics.

[0015] As a preferred embodiment of the above-mentioned first aspect, in the cross-modal contrastive learning training strategy, the SMILES feature vectors, the global graph representation of the two-dimensional molecular graph and the three-dimensional molecular conformation features belonging to the same chemical reaction are used as positive samples, and the SMILES feature vectors, the global graph representation of the two-dimensional molecular graph and the three-dimensional molecular conformation features of different chemical reactions are used as negative samples. The contrast loss between the two different modalities is calculated respectively, and the weighted sum of the calculated contrast losses is taken as the total contrast loss. The network parameters of the microscopic multimodal encoder are updated based on the minimization of the total contrast loss.

[0016] As a preferred embodiment of the first aspect, the adaptive data screening algorithm operates by the following steps:

[0017] S31. First, the microscopic multimodal chemical reaction yield prediction model is trained for several rounds of iterations using the labeled yield dataset to ensure that the microscopic multimodal chemical reaction yield prediction model has preliminary prediction capabilities. During the initial training process, the weights of all training data are set to 1, and a yield data screening vector is constructed to record the weights of each training data.

[0018] S32. After initial training is completed, the reliable data screening phase begins. The microscopic multimodal chemical reaction yield prediction model evaluates the loss function value of each training data point and selects training data with a loss function value less than a preset screening threshold as reliable data. The reliable data is used to update the labeled yield dataset. The weight of the reliable data point is then set to 1, and the weight of the remaining training data point is set to 0.

[0019] S33. After reliable data is screened, the microscopic multimodal chemical reaction yield prediction model is retrained on the updated annotated yield dataset, the screening threshold is increased, and new reliable data is generated and the annotated yield dataset is updated according to step S32;

[0020] S34. Continuously iterate the training according to step S33 until the microscopic multimodal chemical reaction yield prediction model converges to obtain a trained microscopic multimodal chemical reaction yield prediction model.

[0021] As a preferred embodiment of the first aspect, the SMILES sequence encoder adopts a BERT model.

[0022] As a preferred embodiment of the first aspect, the molecular graph encoder adopts a message passing neural network.

[0023] In a second aspect, the present invention provides a computer program product comprising a computer program / instruction, which, when executed by a processor, can implement a multimodal chemical reaction yield prediction method based on adaptive data screening as described in any one of the schemes of the first aspect above.

[0024] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the multimodal chemical reaction yield prediction method based on adaptive data screening as described in any of the schemes of the first aspect above is implemented.

[0025] In a fourth aspect, the present invention provides a computer electronic device comprising a memory and a processor;

[0026] The memory is used to store computer programs;

[0027] The processor is used to implement the multimodal chemical reaction yield prediction method based on adaptive data screening as described in any of the solutions of the first aspect when executing the computer program.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention proposes a multimodal chemical reaction yield prediction method based on adaptive data screening, which can effectively solve the problems of insufficient accuracy and poor robustness of existing models when facing high-noise samples and multimodal data fusion. By introducing an adaptive data screening algorithm constructed by a weakly supervised learning mechanism, the model can dynamically adjust the sample screening strategy during training, give priority to fitting with high-quality samples, effectively resist the interference of low-quality data, and improve training stability. In addition, the present invention realizes the deep fusion of various chemical reaction information such as one-dimensional SMILES sequences, two-dimensional molecular graphs and three-dimensional molecular conformations by constructing a microscopic multimodal encoder, which can extract effective features from different levels and angles, enhance the expression ability of the model and the modeling ability of complex structural relationships. Combined with the cross-modal comparative learning strategy, the present invention further improves the alignment degree of multimodal features in the latent space and accelerates the convergence efficiency of the model. At the same time, the uncertainty quantification module integrated in the present invention can estimate the credibility of the prediction results, provide decision support for practical applications, and enhance the interpretability and controllability of the model. Experimental verification has shown that the method of the present invention achieves better prediction performance than existing models on multiple standard reaction datasets (including Buchwald HTE, Suzuki HTE and USPTO), showing higher accuracy, robustness and generalization ability, and is suitable for various chemical reaction prediction scenarios such as drug synthesis and high-throughput experimental screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of the steps of the present invention;

[0031] Figure 2 The overall architecture of the microscopic multimodal chemical reaction yield prediction model provided in this embodiment;

[0032] Figure 3 Schematic diagram of the adaptive data screening algorithm provided in this embodiment;

[0033] Figure 4 Schematic diagram of the process of pre-training the microscopic multimodal encoder in this embodiment. DETAILED DESCRIPTION

[0034] In order to make the above-mentioned objects, features and advantages of the present invention more clearly understood, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in the various embodiments of the present invention can be combined accordingly without conflicting with each other.

[0035] Before further explaining the embodiments of the present invention in detail, the nouns and terms involved in the embodiments of the present invention are explained. The nouns and terms involved in the embodiments of the present invention are subject to the following interpretations:

[0036] 1) Chemical reaction yield

[0037] The yield of a chemical reaction refers to the efficiency with which reactants are converted into products in a chemical reaction. It is usually represented by the yield value y, which ranges from 0 to 1. When y = 0, the reaction cannot occur; when y = 1, the reaction conversion rate reaches 100%.

[0038] 2) One-dimensional SMILES sequence

[0039] SMILES (Simplified Molecular Input Line Entry System) is a string format used to represent chemical structures. Using SMILES, chemical structures can be represented in a concise text format that can be parsed and processed by computer programs. In this paper, SMILES sequences are used to describe reactants and products in chemical reactions.

[0040] 3) Two-dimensional molecular graph

[0041] A two-dimensional molecular graph is a graphical representation that uses vertex and edge relationships to show the connectivity and local structure of molecules. In this graph, each node represents an atom, and the edges between nodes represent the chemical bonds between atoms. A two-dimensional molecular graph can provide information about the overall structure and local environment of a molecule.

[0042] 4) Three-dimensional molecular conformation

[0043] A 3D molecular conformation refers to the positional coordinates of atoms in a molecule in three-dimensional space and their stereological relationships. This 3D structural information provides more detailed spatial arrangement and structural details than a 2D molecular diagram, and is crucial for understanding the physical and chemical properties of molecules and modeling intermolecular interactions.

[0044] 5) Adaptive data screening

[0045] Adaptive data filtering is a mechanism that dynamically adjusts the filtering threshold to prioritize high-quality labeled samples and mitigate the effects of data noise. In this paper, the adaptive data filter mimics human learning strategies, processing information from the shallow end to the deep end. This allows the model to initially focus on samples with strong regularity and ease of fitting, gradually transitioning to more complex samples as the model's capabilities improve.

[0046] 6) Multimodal feature extraction

[0047] Multimodal feature extraction refers to the process of extracting feature information from data of multiple modalities. In this paper, the multimodal feature extractor can integrate one-dimensional SMILES sequences, two-dimensional molecular graphs, and three-dimensional molecular conformation information, promoting the model to deeply understand chemical information from multiple perspectives and levels.

[0048] 7) Cross-modal contrastive learning

[0049] Cross-modal contrastive learning is a method that accelerates model learning by comparing the similarities and differences between feature vectors from different modalities. In this paper, a cross-modal contrastive learning strategy is used to align and fuse feature vectors from different modalities, thereby accelerating the model learning process.

[0050] 8) Uncertainty Quantification

[0051] Uncertainty quantification refers to the process of evaluating the credibility of model predictions. In this paper, the uncertainty quantification module is used to evaluate the credibility of model predictions, provide uncertainty estimates for the model's predictions, and help users better understand the model's prediction quality and reliability.

[0052] 9) Weakly supervised learning

[0053] Weakly supervised learning is a machine learning method in which the model's training data contains only partial or inaccurate label information. In this paper, weakly supervised learning is used to develop an adaptive yield data screening algorithm, aiming to prioritize training the model to fit high-quality labeled samples, thereby effectively resisting the interference of data noise.

[0054] 10) Micro-multimodal features

[0055] Microscopic multimodal features refer to multiple modal feature information extracted at the microscopic level, including one-dimensional sequence information, two-dimensional topological information, and three-dimensional conformational information. In this paper, microscopic multimodal features are integrated through a multimodal feature extractor to facilitate the model's in-depth understanding of chemical information from multiple perspectives and levels.

[0056] Based on the above explanations of the nouns and terms involved in the embodiments of this application, the following will first explain the implementation scenario of the present invention. In view of the technical defects of the existing yield prediction model being limited by data label noise and uneven distribution, the present invention innovatively proposes a solution that combines adaptive progressive learning with multimodal feature fusion, namely a multimodal chemical reaction yield prediction method based on adaptive data screening. Figure 1 As shown, in a preferred implementation of the present invention, the multimodal chemical reaction yield prediction method based on adaptive data screening includes the following steps S1 to S4. The specific implementation process is described below.

[0057] S1. Obtain a one-dimensional SMILES sequence and its corresponding two-dimensional molecular graph to form an unlabeled multimodal dataset.

[0058] S2. In the pre-training phase, the network parameters of the multi-layer perceptron predictor are fixed, and the micro-multimodal encoder is trained on an unlabeled multimodal dataset using a cross-modal contrastive learning training strategy.

[0059] It should be noted that in the microscopic multimodal encoder of step S2 of the present invention, the one-dimensional SMILES sequence is encoded by the SMILES sequence encoder, and the one-dimensional SMILES sequence is mapped to the vector space to obtain a SMILES feature vector, and the two-dimensional molecular graph is encoded by the molecular graph encoder to generate a global graph representation of the two-dimensional molecular graph, and a three-dimensional molecular conformation feature is generated based on the two-dimensional molecular graph. Then, the SMILES feature vector, the global graph representation of the two-dimensional molecular graph and the three-dimensional molecular conformation feature are combined through a feature splicing operation to generate a comprehensive feature vector.

[0060] In this embodiment, the SMILES sequence encoder adopts the BERT model. The model uses its bidirectional attention mechanism to simultaneously consider the contextual information on the left and right sides of each element in the SMILES sequence, thereby effectively capturing the complex relationships of the molecular structure.

[0061] In this embodiment, the molecular graph encoder uses a message passing neural network (MPNN), which aggregates local graph structure information by iteratively passing messages between nodes, ultimately generating a global graph representation. MPNN is capable of capturing the complex relationships between atoms in a molecule, including both long-range and short-range interactions.

[0062] It should be noted that, in the present invention, for the three-dimensional molecular conformational features, they can be encoded by an encoder, or the features can be directly extracted by calculation. In the present embodiment, the three-dimensional molecular conformational features are extracted by calculating a series of three-dimensional molecular fingerprints. Specifically, the generation method of the above-mentioned three-dimensional molecular conformational features is as follows: first, the three-dimensional molecular initial conformation is obtained from the two-dimensional molecular graph using a technology related to conformational prediction, and the energy of the three-dimensional molecular initial conformation is minimized by the MMFF molecular mechanics force field to obtain the corresponding stable conformation and form a series of three-dimensional molecular stable conformation sets participating in chemical reactions; the chemical informatics and machine learning toolkit RDKit is then used to calculate the stable conformation set to obtain three-dimensional molecular conformational features including geometric diameter, moment of inertia, and mass distribution, or the three-dimensional molecular conformation encoder is used to encode the stable conformation set to generate three-dimensional molecular conformational features to enrich the feature representation of the molecule.

[0063] In this embodiment, if Figure 4As shown, the present application constructs a micro multi-modal encoder, which can integrate one-dimensional SMILES sequence, two-dimensional molecular graph and three-dimensional molecular conformation. Specifically, the SMILES sequence is encoded by the SMILES sequence encoder (Transformer encoder in Figure 4 ), the two-dimensional molecular graph is encoded by the molecular graph encoder (MPNN encoder in Figure 4 ), and the three-dimensional molecular conformation feature is obtained by encoding the three-dimensional molecular stable conformation by the SchNet encoder. Finally, the comprehensive feature vector (multi-modal fusion feature in Figure 4 ) is generated by feature splicing operation, which is used as the input of the multi-layer perceptron predictor to facilitate the model to deeply understand the chemical information from multiple angles and levels.

[0064] It should be noted that in the cross-modal contrastive learning training strategy, the SMILES feature vector, the global graph representation of the two-dimensional molecular graph and the three-dimensional molecular conformation feature belonging to the same chemical reaction are taken as positive samples, and the SMILES feature vector, the global graph representation of the two-dimensional molecular graph and the three-dimensional molecular conformation feature of different chemical reactions are taken as negative samples. The contrastive loss between the two different modalities is calculated, and the weighted sum of the calculated contrastive losses is taken as the total contrastive loss, and the network parameters of the micro multi-modal encoder are updated based on the minimum total contrastive loss.

[0065] In this embodiment, in order to facilitate the fusion of model knowledge, a cross-modal contrastive learning training strategy is introduced. This strategy is based on the idea that the three feature vectors encoded from the same chemical reaction should be pulled closer to each other as positive samples, and the three feature vectors encoded from different chemical reactions should be pulled away from each other as negative samples. Through this training strategy, the model can align the feature vectors from different modalities in the hidden space, thereby accelerating the learning process. Specifically, a nonlinear projection function g(·) is obtained, which is used to map the multi-modal encoding vector to a fixed-dimensional vector for contrastive learning. Then the function is used to map the SMILES feature vector Global graph representation of two-dimensional molecular graph And three-dimensional molecular conformation feature To obtain the encoding vector obtained from the SMILES feature vector The encoding vector obtained from the global graph representation of the two-dimensional molecular graph And the encoding vector obtained from the three-dimensional molecular conformation feature Then, h S , h G and h C are each regarded as a modality, and the contrastive loss between any two modalities is first calculated

[0066]

[0067] where h1, h2 represent h S , h G and h C ; N represents the number of chemical reactions; represents the encoding vector of the first modality corresponding to the jth chemical reaction; represents the encoding vector of the second modality corresponding to the jth chemical reaction; represents the encoding vector of the first modality corresponding to the kth chemical reaction; represents the encoding vector of the second modality corresponding to the kth chemical reaction; j and k are the chemical reaction indices; <·,·> is the vector inner product, used to represent the similarity of two encoding vectors; τ is the temperature parameter, used to control the entropy of the distribution.

[0068] Thus, the total contrastive loss of the three modalities can be expressed as:

[0069]

[0070] S3. After pre-training, the unlabeled multi-modal data set is labeled to obtain a labeled yield data set, the network parameters of the multi-layer perceptron predictor are unfrozen, a micro multi-modal chemical reaction yield prediction model is constructed by the micro multi-modal encoder and the multi-layer perceptron predictor, the micro multi-modal chemical reaction yield prediction model is end-to-end jointly trained based on the labeled yield data set, the micro multi-modal chemical reaction yield prediction model weights are synchronously updated through back propagation, and a trained micro multi-modal chemical reaction yield prediction model is obtained;

[0071] In the joint training process, the adaptive data screening algorithm dynamically removes abnormal training data according to the prediction confidence and feature space distribution, iteratively updates the labeled yield data set, calculates the variance estimate of the chemical reaction yield prediction value through the uncertainty quantification module, and generates the confidence interval of the chemical reaction yield prediction value based on Monte Carlo sampling.

[0072] It should be noted that, as Figure 2As shown, the total architecture diagram of the micro multi-modal chemical reaction yield prediction model provided by the embodiment of the present application is provided with an adaptive data screening mechanism. The model combines an adaptive data screening algorithm based on weakly supervised learning, and contains two parts of a micro multi-modal encoder and a multi-layer perception predictor. Among them, the adaptive data screening algorithm preferentially screens high-quality labeled training data by dynamically adjusting the screening threshold to resist the interference of data noise; the micro multi-modal encoder can integrate one-dimensional SMILES sequence, two-dimensional molecular graph and three-dimensional molecular conformation, and promote the model to deeply understand chemical information from multiple angles and levels; the multi-layer perception predictor is used to generate the final chemical reaction yield prediction value.

[0073] The adaptive data screening algorithm designed by the present application will be briefly described below. As shown in the figure, the algorithm preferentially screens high-quality labeled samples by dynamically adjusting the screening threshold, so that the model pays attention to samples with stronger regularity and easier to fit in the early training stage, and gradually transits to processing more complex samples as the model capacity improves. This process is similar to the process of human learning complex knowledge, that is, by gradually increasing the difficulty to improve the learning efficiency. Figure 3

[0074] In specific implementation, the one-dimensional SMILES sequence and its corresponding two-dimensional molecular graph are used as training data, and the adaptive data screening algorithm operates through the following steps:

[0075] S31. Model initialization: first, use the annotated yield data set to initialize the micro multi-modal chemical reaction yield prediction model for several rounds of iterative training, so that the micro multi-modal chemical reaction yield prediction model has a preliminary prediction ability (i.e. the micro multi-modal chemical reaction yield prediction model can initially learn some basic reaction representation and yield prediction ability). During the initialization training process, the weight of all training data is set to 1, and a yield data screening vector is constructed to record the weight of each training data.

[0076] S32. Reliable data screening: after the initialization training is completed, enter the reliable data screening stage, and the micro multi-modal chemical reaction yield prediction model screens out the training data with a loss function value less than the preset screening threshold by evaluating the loss function value of each training data and uses it as reliable data. Update the annotated yield data set with reliable data, and set the weight of reliable data to 1 and the weight of the rest of the training data to 0.

[0077] In this embodiment, these reliable data update the original annotated yield data set and are used for subsequent model training to update the model parameters and improve the prediction performance of the model.

[0078] ​S33. Dynamic adjustment of threshold: After reliable data screening, retrain the micro multi-modal chemical reaction yield prediction model on the updated labeled yield dataset, increase the screening threshold, and generate new reliable data according to step S32 to update the labeled yield dataset.

[0079] In this embodiment, as the training progresses, the ability of the micro multi-modal chemical reaction yield prediction model gradually improves, and the screening threshold is dynamically adjusted. In the initial stage, the screening threshold is low to ensure that the model can screen reliable data with high enough confidence; as the model performance improves, the screening threshold is gradually increased, and more challenging training data is added to improve the generalization ability of the model, until the model converges.

[0080] It should be noted that the present application also integrates an uncertainty quantification module for evaluating the reliability of the model's predictions. This module can provide uncertainty estimates for the model's prediction results, helping users better understand the quality and reliability of the model's predictions, and thus making more informed decisions in practical applications. The implementation of the uncertainty quantification module is a prior art, in order to facilitate the understanding of the present application by those skilled in the art, the module will be briefly described.

[0081] In this embodiment, specifically, the uncertainty quantification module operates through the following steps:

[0082] 1. Prediction result evaluation: For each prediction result, the uncertainty quantification module calculates the variance of its predicted value as an indicator of uncertainty estimation.

[0083] 2. Uncertainty threshold setting: According to the requirements of the application scenario, set the uncertainty threshold. When the uncertainty of the prediction result exceeds the threshold, the model will prompt the user that the reliability of the prediction result is low and needs to be further verified.

[0084] 3. Result feedback and optimization: Users can further verify and optimize the prediction results according to the prompts of the uncertainty quantification module, thereby improving the prediction performance of the model.

[0085] S4. Input the one-dimensional SMILES sequence and its corresponding two-dimensional molecular graph to be predicted into the trained micro multi-modal chemical reaction yield prediction model, extract high-dimensional fusion features by the micro multi-modal encoder, and then map the extracted high-dimensional fusion features to the chemical reaction yield prediction results by the multi-layer perception predictor. Calculate the variance estimate of the chemical reaction yield prediction result by the uncertainty quantification module, generate the confidence interval of the chemical reaction yield prediction result based on Monte Carlo sampling, and realize multi-modal chemical reaction yield prediction based on adaptive data screening.

[0086] In the present embodiment, the micro multi-modal encoder integrates multi-dimensional molecular features into a comprehensive feature vector (i.e., dimension fusion feature) through a multi-modal feature fusion strategy. Specifically, for the SMILES feature vector Global graph representation of two-dimensional molecular graph And three-dimensional molecular conformation features The present application combines a feature fusion function v(·) with And to achieve the purpose of fusing feature information of different modalities, generating high-dimensional fusion features

[0087]

[0088] Wherein, represents a feature splicing operation. The generated high-dimensional fusion features are input into a multi-layer perceptron predictor to facilitate the model to deeply understand chemical information from multiple angles and levels, and finally output a chemical reaction yield prediction value by processing the multi-layer perceptron predictor.

[0089] Through the above technical solutions, the present application can effectively resist the interference of data noise, deeply fuse multi-modal information, and thus improve the accuracy and reliability of chemical reaction yield prediction.

[0090] The present application will now demonstrate the application effect of the multi-modal chemical reaction yield prediction method based on adaptive data screening described in S1-S4 of the above embodiment on specific data sets through a specific example, so as to facilitate understanding of the essence of the present application.

[0091] Embodiment

[0092] The specific implementation process of the multi-modal chemical reaction yield prediction method based on adaptive data screening used in the present embodiment is as described above, and will not be repeated here.

[0093] In order to verify the effectiveness of the weakly supervised adaptive training strategy and multi-modal fusion mechanism proposed by the present application in the chemical reaction yield prediction task, the present embodiment carries out comparative experiments on Buchwald HTE, Suzuki HTE and USPTO data sets. The data sets used cover representative chemical reaction scenarios of different reaction types, experimental scales and noise levels. The comparative models include the current mainstream sequence modeling method YieldBert, the fingerprint feature method DRFP based on molecular structure changes, the graph neural network model AGNN with attention mechanism, the MPNN model based on message passing mechanism, and the Egret method using conditional contrast learning strategy to improve the robustness of the model. These models are representative in structure design and are widely used in yield prediction tasks, and have strong reference value. The experimental results are shown in Tables 1-3.

[0094] Table 1 Performance evaluation of the present application on Buchwald HTE dataset

[0095] index YieldBert DRFP AGNN MPNN Egret Method of the present invention MAE 3.91 3.90 3.89 2.92 4.01 2.90 RMSE 6.02 6.03 6.01 4.43 6.19 4.35 <![CDATA[R 2 ]]> 0.951 0.950 0.953 0.974 0.940 0.977

[0096] Table 1 shows the yield prediction performance of different models on Buchwald HTE dataset. This dataset contains 3955 palladium-catalyzed Buchwald-Hartwig C-N coupling reactions, involving 15 aryl halides, 4 ligands, 3 bases and 23 additive combinations, belonging to typical high-throughput experiment (HTE) data. The reaction types are relatively uniform, and the similarity between molecular structures is high. The results in Table 1 show that the present application method is superior to the existing baseline methods in R 2 , MAE and RMSE, especially in the coefficient of determination R 2 up to 0.977, indicating that it has stronger feature extraction ability and fitting accuracy in such low-heterogeneity data.

[0097] Table 2 Performance evaluation of the present application on Suzuki HTE dataset

[0098] index YieldBert DRFP AGNN MPNN Egret Method of the present invention MAE 8.13 6.77 6.96 6.02 6.76 6.03 RMSE 12.07 10.59 11.01 9.41 10.59 9.22 [R 2 ]]> 0.810 0.850 0.840 0.886 0.850 0.889

[0099] Table 2 shows the prediction results of different models on Suzuki HTE dataset. This dataset contains 5760 Suzuki-Miyaura C-C coupling reactions, composed of 15 electrophile / nucleophile combinations, 12 ligands, 8 bases and 4 solvents, with more complex and diverse reaction conditions, stronger data heterogeneity and higher noise level. The results in Table 2 show that the present application method still achieves the optimal R 2 (0.889) and RMSE value on this dataset, indicating that the proposed model has good generalization ability and noise resistance, and still performs stably under complex condition combinations.

[0100] Table 3 Performance evaluation of the present application on USPTO

[0101]

[0102] Table 3 is the yield prediction results of two subsets of the USPTO real patent data set, USPTO-gram (gram-scale reactions) and USPTO-subgram (milligram-scale reactions) respectively. The data set is derived from the US patent publication, covering various types of actual reactions such as C-C coupling, C-N coupling, oxidation-reduction reaction, etc., with significant scale unevenness, large noise, and multiple reaction types. Table 3 shows the prediction performance of the model on the two subsets. From the results, it can be seen that the method still maintains good stability on actual large-scale complex reaction data, and is superior to existing models in multiple indicators, reflecting strong generalization ability and practical applicability.

[0103] In summary, from Tables 1-3, it can be seen that the method exhibits significant advantages on data sets of different reaction types, different scales, and different distribution characteristics, with stronger prediction accuracy, robustness, and practical application potential.

[0104] It can be understood that the multi-modal chemical reaction yield prediction method based on adaptive data screening described in S1-S4 can be realized by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present application also provides a computer program product corresponding to the multi-modal chemical reaction yield prediction method based on adaptive data screening provided by the above-mentioned embodiments, which includes computer programs / instructions that can be executed by a processor to realize the multi-modal chemical reaction yield prediction method based on adaptive data screening as described in the above-mentioned embodiments.

[0105] Similarly, based on the same inventive concept, another preferred embodiment of the present application also provides a computer electronic device corresponding to the multi-modal chemical reaction yield prediction method based on adaptive data screening provided by the above-mentioned embodiments, which includes a memory and a processor;

[0106] The memory is used to store computer programs;

[0107] The processor is used to realize the multi-modal chemical reaction yield prediction method based on adaptive data screening in the above-mentioned embodiments when executing the computer programs.

[0108] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of a software function unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application.

[0109] Therefore, based on the same inventive concept, another preferred embodiment of the present application also provides a computer readable storage medium corresponding to the multi-modal chemical reaction yield prediction method based on adaptive data filtering provided by the above-mentioned embodiments. The storage medium stores a computer program, and when the computer program is executed by a processor, the multi-modal chemical reaction yield prediction method based on adaptive data filtering in the above-mentioned embodiments can be realized.

[0110] It can be understood that the above-mentioned storage medium can include a random access memory (RAM) and can also include a non-volatile memory (NVM), such as at least one disk memory. Meanwhile, the storage medium can also be a U disk, a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0111] It can be understood that the above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0112] In addition, it should be noted that, for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the foregoing method embodiments, and will not be described here. In the embodiments provided in the present application, the division of steps or modules in the system and method described is only a logical functional division, and there can be another division manner in actual implementation, for example, multiple modules or steps can be combined or integrated together, or a module or step can be split.

[0113] The above-described embodiments are only a preferred scheme of the present application, and are not intended to limit the present application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application. Therefore, any technical scheme obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present application.

Claims

1. A multimodal chemical reaction yield prediction method based on adaptive data screening, characterized in that: The following steps are involved: S1. Obtain a one-dimensional SMILES sequence and its corresponding two-dimensional molecular graph to form an unlabeled multimodal dataset; S2. In the pre-training phase, the network parameters of the multi-layer perceptron predictor are fixed, and the micro-multimodal encoder is trained on the unlabeled multimodal dataset using a cross-modal contrastive learning training strategy. S3. After pre-training is complete, the unlabeled multimodal dataset is labeled to obtain a labeled yield dataset. The network parameters of the multi-layer perceptron predictor are unfrozen, and a microscopic multimodal chemical reaction yield prediction model is constructed from the microscopic multimodal encoder and the multi-layer perceptron predictor. The microscopic multimodal chemical reaction yield prediction model is end-to-end jointly trained based on the labeled yield dataset. The weights of the microscopic multimodal chemical reaction yield prediction model are synchronously updated through backpropagation to obtain a trained microscopic multimodal chemical reaction yield prediction model. During the joint training process, an adaptive data screening algorithm dynamically removes abnormal training data based on prediction confidence and feature space distribution, iteratively updates the labeled yield dataset, and calculates the variance estimate of the chemical reaction yield prediction value through the uncertainty quantification module. Confidence intervals for the chemical reaction yield prediction value are generated based on Monte Carlo sampling. S4. The one-dimensional SMILES sequence to be used for yield prediction and its corresponding two-dimensional molecular graph are input into the trained microscopic multimodal chemical reaction yield prediction model. The microscopic multimodal encoder extracts high-dimensional fusion features, and the multi-layer perceptron predictor maps the extracted high-dimensional fusion features into chemical reaction yield prediction results. The uncertainty quantification module calculates the variance estimate of the chemical reaction yield prediction results, and the confidence interval of the chemical reaction yield prediction results is generated based on Monte Carlo sampling, thereby realizing multimodal chemical reaction yield prediction based on adaptive data screening.

2. A multimodal chemical reaction yield prediction method based on adaptive data screening according to claim 1, characterized in that: In the microscopic multimodal encoder of step S2, the one-dimensional SMILES sequence is encoded by the SMILES sequence encoder, and the one-dimensional SMILES sequence is mapped to the vector space to obtain a SMILES feature vector. The two-dimensional molecular graph is encoded by the molecular graph encoder to generate a global graph representation of the two-dimensional molecular graph. Based on the two-dimensional molecular graph, three-dimensional molecular conformation features are generated. Then, the SMILES feature vector, the global graph representation of the two-dimensional molecular graph, and the three-dimensional molecular conformation features are combined through feature splicing operations to generate a comprehensive feature vector.

3. A multimodal chemical reaction yield prediction method based on adaptive data screening according to claim 2, characterized in that: The specific method for generating the three-dimensional molecular conformational features is as follows: first, a technology related to conformation prediction is used to obtain the initial conformation of the three-dimensional molecule from the two-dimensional molecular graph, and the energy of the initial conformation of the three-dimensional molecule is minimized through the MMFF molecular mechanics force field to obtain the corresponding stable conformation and form a series of three-dimensional molecular stable conformation sets participating in chemical reactions; then, the chemical informatics and machine learning toolkit RDKit is used to calculate the stable conformation set to obtain the three-dimensional molecular conformational features including geometric diameter, moment of inertia, and mass distribution, or the three-dimensional molecular conformation encoder is used to encode the stable conformation set to generate the three-dimensional molecular conformational features.

4. A multimodal chemical reaction yield prediction method based on adaptive data screening according to claim 2, characterized in that: In the cross-modal contrastive learning training strategy, the SMILES feature vectors, the global graph representation of the two-dimensional molecular graph, and the three-dimensional molecular conformational features belonging to the same chemical reaction are used as positive samples, and the SMILES feature vectors, the global graph representation of the two-dimensional molecular graph, and the three-dimensional molecular conformational features of different chemical reactions are used as negative samples. The contrast loss between the two different modalities is calculated respectively, and the weighted sum of the calculated contrast losses is taken as the total contrast loss. The network parameters of the microscopic multimodal encoder are updated based on the minimization of the total contrast loss.

5. The multimodal chemical reaction yield prediction method based on adaptive data screening according to claim 1, characterized in that: The adaptive data filtering algorithm operates through the following steps: S31. First, the microscopic multimodal chemical reaction yield prediction model is trained for several rounds of iterations using the labeled yield dataset to ensure that the microscopic multimodal chemical reaction yield prediction model has preliminary prediction capabilities. During the initial training process, the weights of all training data are set to 1, and a yield data screening vector is constructed to record the weights of each training data. S32. After initial training is completed, the reliable data screening phase begins. The microscopic multimodal chemical reaction yield prediction model evaluates the loss function value of each training data point and selects training data with a loss function value less than a preset screening threshold as reliable data. The reliable data is used to update the labeled yield dataset. The weight of the reliable data point is then set to 1, and the weight of the remaining training data point is set to 0. S33. After reliable data is screened, the microscopic multimodal chemical reaction yield prediction model is retrained on the updated annotated yield dataset, the screening threshold is increased, and new reliable data is generated and the annotated yield dataset is updated according to step S32; S34. Continuously iterate the training according to step S33 until the microscopic multimodal chemical reaction yield prediction model converges to obtain a trained microscopic multimodal chemical reaction yield prediction model.

6. A multimodal chemical reaction yield prediction method based on adaptive data screening according to claim 2, characterized in that: The SMILES sequence encoder adopts the BERT model.

7. A multimodal chemical reaction yield prediction method based on adaptive data screening according to claim 2, characterized in that: The molecular graph encoder adopts a message passing neural network.

8. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, it can implement the multimodal chemical reaction yield prediction method based on adaptive data screening as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by the processor, the multimodal chemical reaction yield prediction method based on adaptive data screening according to any one of claims 1 to 7 is implemented.

10. A computer electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is configured to implement the multimodal chemical reaction yield prediction method based on adaptive data screening as described in any one of claims 1 to 7 when executing the computer program.

Images