Data processing method, device, model training method, device and electronic device

Abstract

This application relates to a data processing method, apparatus, and model training method, apparatus, and electronic device. The data processing method includes: obtaining first features of m training data points in a training dataset, where m is an integer greater than or equal to 1; obtaining n target data points with second features, so as to adjust the model parameters of a pre-trained model based on at least a portion of the n target data points, where n is an integer greater than or equal to 1; wherein the first and second features satisfy a similarity requirement, the pre-trained model is trained based on the training dataset, and the target data points may be unlabeled or labeled. This application can improve the applicability of training data, thereby improving the model training effect.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence and computer simulation technology, and in particular to a data processing method, device, model training method, device and electronic equipment. Background Art

[0002] With the rapid development of computer technology and artificial intelligence technology, computer simulation technology is being applied to more and more scenarios, such as material design, drug design, etc.

[0003] However, the applicant found that the prediction accuracy of the pre-trained model obtained by the relevant technology needs to be improved when applied to a certain application domain. Summary of the invention

[0004] In order to solve or partially solve the problems existing in the related art, the present application provides a data processing method, device and model training method, device and electronic device, which can effectively improve the accuracy of the pre-trained model in a specific application domain.

[0005] The first aspect of the present application provides a data processing method, comprising: obtaining a first feature of m training data in a training data set, where m is an integer greater than or equal to 1; obtaining n target data with a second feature, so as to adjust a model parameter of a pre-trained model based on at least part of the n target data, where n is an integer greater than or equal to 1; wherein the first feature and the second feature meet a similarity requirement, the pre-trained model is trained based on the training data set, and the target data has no label information or has label information.

[0006] The second aspect of the present application provides a model training method, comprising: obtaining a first target data set, the first target data set comprising target data obtained according to the above method, and the target data has no label information; using the first target data set to train a pre-trained model to adjust the model parameters of the pre-trained model to obtain an adjusted pre-trained model.

[0007] The third aspect of the present application provides a method for predicting molecular properties, comprising: obtaining molecular data; based on the molecular data, predicting the molecular properties of the molecular data using a molecular prediction model obtained by the adjusted pre-trained model as described above.

[0008] The fourth aspect of the present application provides a method for evaluating molecular properties, comprising: obtaining molecular data, the molecular data having label information; based on the molecular data, evaluating using a molecular evaluation model obtained by the adjusted pre-trained model as described above to obtain a molecular evaluation result of the molecular data.

[0009] The fifth aspect of the present application provides a design method, comprising: obtaining molecular properties and / or molecular evaluation results according to the above method; and performing drug design or material design based on the molecular properties and / or molecular evaluation results.

[0010] The sixth aspect of the present application provides a data processing device, including: a first feature acquisition module, used to obtain the first feature of m training data in a training data set, where m is an integer greater than or equal to 1; a target data acquisition module, used to obtain n target data with a second feature, so as to adjust the model parameters of a pre-trained model based on at least part of the n target data, where n is an integer greater than or equal to 1; wherein the first feature and the second feature meet the similarity requirement, the pre-trained model is trained based on the training data set, and the target data has no label information or has label information.

[0011] The seventh aspect of the present application provides a model training device, including: a data set acquisition module, used to obtain a first target data set, the first target data set includes target data obtained according to the above-mentioned device, and the target data has no label information; a model training module, used to use the first target data set to train a pre-trained model to adjust the model parameters of the pre-trained model to obtain an adjusted pre-trained model.

[0012] The eighth aspect of the present application provides a molecular property prediction device, including: a first molecular data acquisition module, used to obtain molecular data; a molecular property prediction module, used to predict the molecular properties of the molecular data using a molecular prediction model obtained according to the above-mentioned adjusted pre-trained model.

[0013] The ninth aspect of the present application provides a molecular property evaluation device, comprising: a second molecular data acquisition module, used to obtain molecular data; a molecular property prediction module, used to evaluate using a molecular evaluation model obtained according to the above-mentioned adjusted pre-trained model to obtain a molecular evaluation result of the molecular data.

[0014] The tenth aspect of the present application provides a design device, which includes: an information acquisition module, used to obtain molecular properties according to the above-mentioned molecular property prediction device, and / or obtain molecular evaluation results according to the above-mentioned molecular property evaluation device; a design module, used to perform drug design or material design based on molecular properties and / or molecular evaluation results.

[0015] The eleventh aspect of the present application provides an electronic device, comprising: a processor; and a memory on which executable codes are stored, and when the executable codes are executed by the processor, the processor executes the method.

[0016] The twelfth aspect of the present application also provides a computer-readable storage medium on which executable code is stored. When the executable code is executed by a processor of an electronic device, the processor executes the above method.

[0017] The thirteenth aspect of the present application also provides a computer program product, including executable code, which implements the above method when executed by a processor.

[0018] The data processing method, device, model training method, device and electronic device provided in the present application obtain target data similar to the training data in the training data set for a pre-trained model trained by a training data set, so that the pre-trained model can be adaptively adjusted based on the target data, which helps to improve the prediction accuracy of the pre-trained model.

[0019] In addition, the application domain of the target data is the same as that of the training data set. A small-scale pre-training is performed on the application domain data, and then fine-tuning is performed on the downstream tasks for the actual supervision tasks, which can effectively improve the prediction accuracy of the pre-training model.

[0020] In addition, pre-training for application domain data can be unsupervised training, which does not require data labeling. It will not incur additional labor costs while improving the prediction accuracy of the model.

[0021] In addition, target data can be obtained based on Morgan fingerprint and / or maximum molecular structure search to improve the quality of the obtained training data and improve the model fine-tuning effect.

[0022] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above and other objects, features and advantages of the present application will become more apparent by describing in more detail exemplary embodiments of the present application in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the present application.

[0024] Figure 1 An exemplary system architecture to which a data processing method, device, model training method, device, and electronic device according to an embodiment of the present application can be applied is schematically shown;

[0025] Figure 2 The flowchart of the data processing method according to the embodiment of the present application is schematically shown;

[0026] Figure 3 A schematic diagram schematically shows a first fingerprint feature according to an embodiment of the present application;

[0027] Figure 4 A schematic diagram schematically shows a molecular structure according to an embodiment of the present application;

[0028] Figure 5 A schematic diagram of a process for obtaining target training data according to an embodiment of the present application is schematically shown;

[0029] Figure 6 A schematic diagram of another process of obtaining target training data according to an embodiment of the present application is schematically shown;

[0030] Figure 7 A schematic diagram of a process for determining a maximum common substructure according to an embodiment of the present application is schematically shown;

[0031] Figure 8 A flowchart of a model training method according to an embodiment of the present application is schematically shown;

[0032] Fig. 9 The following schematically shows an architecture diagram of a model training system according to an embodiment of the present application;

[0033] Fig.10 A schematic diagram schematically shows a model training process according to an embodiment of the present application;

[0034] Fig.11 A flowchart of a method for predicting molecular properties according to an embodiment of the present application is schematically shown;

[0035] Fig.12 A flowchart of a molecular property evaluation method according to an embodiment of the present application is schematically shown;

[0036] Fig.13 A flowchart of a design method according to an embodiment of the present application is schematically shown;

[0037] Fig.14 A block diagram of a data processing device according to an embodiment of the present application is schematically shown;

[0038] Fig.15 A block diagram of a training device according to an embodiment of the present application is schematically shown;

[0039] Fig.16 A block diagram of a prediction device according to an embodiment of the present application is schematically shown;

[0040] Fig.17 A block diagram of an evaluation device according to an embodiment of the present application is schematically shown;

[0041] Fig.18 A block diagram of a design device according to an embodiment of the present application is schematically shown;

[0042] Fig.19 A block diagram of an electronic device according to an embodiment of the present application is schematically shown. DETAILED DESCRIPTION

[0043] The embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to make the present application more thorough and complete, and to fully convey the scope of the present application to those skilled in the art.

[0044] The terms used in this application are only for the purpose of describing specific embodiments and are not intended to limit this application. The terms "comprise", "include", etc. used herein indicate the existence of features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0045] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification, and should not be interpreted in an idealized or overly rigid manner.

[0046] It should be understood that although the terms "first", "second", "third", etc. may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of this application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0047] Before describing the technical solution of the present application, some technical terms in the field involved in the present application are first explained.

[0048] Molecular descriptors represent molecules as data structures that can be processed by computer programs.

[0049] Molecular graph is a molecular descriptor that represents atoms as nodes and the relationships between atoms as edges.

[0050] The pre-trained model is a model obtained by performing pre-unsupervised training on a large amount of unlabeled data on a pre-set task. The model will then be used for the migration of downstream tasks to achieve better results.

[0051] Model evaluation indicators are used to evaluate the performance of the trained model. Model evaluation indicators include but are not limited to at least one of the following: evaluation index determination coefficient, area under the receiver operating characteristic curve (ROC curve), accuracy, recall, sensitivity, specificity, root mean square error (STD), and mean absolute error.

[0052] Traditional pharmaceutical R&D is characterized by long processes and high failure rates. For example, nearly three-quarters of pre-research drugs fail due to efficacy or safety reasons. Compared with traditional methods such as computer-aided drug design, artificial intelligence-driven drug development (AIDD) methods have gradually been recognized and applied in recent years. Currently, most of the artificial intelligence (AI) solutions in the industry for business needs such as molecular property prediction are based on traditional machine learning or earlier graph neural network technology, and have not taken advantage of deep learning big data methods.

[0053] In this context, data has become a bottleneck for the development of AIDD. Although a certain application domain has a large amount of data without label information (such as data with molecular sequence or structure but no properties, etc.), there is very little data with label information.

[0054] By training the pre-trained model and migrating it to downstream tasks, a trained model with high prediction accuracy can be obtained with only a small amount of training data with labeled information. Pre-trained models have been applied in many industrial scenarios, such as natural language processing and computer vision.

[0055] During the drug development process, the applicant discovered that when the pre-trained model is migrated to the downstream, the input data of the pre-trained model may have problems such as small data volume, high noise, and batch effects between data, which results in the prediction accuracy of the pre-trained model migrated to the downstream being unable to meet user needs. Due to the low accuracy of molecular property prediction based on artificial intelligence models, it cannot meet the screening and design requirements of auxiliary drugs, etc.

[0056] For example, pre-training drug molecules under specific representations and using the pre-trained model for downstream small sample size tasks is a possible effective solution for this scenario. In the field of AI-based drug research and development, a large number of pre-trained models have emerged, including Pretrained-GNN, GROVER, and so on. Related technologies can migrate pre-trained models by increasing the number of network layers in downstream tasks, adjusting only the subnet, etc., but the model prediction accuracy is still difficult to meet user needs.

[0057] There have been no similar attempts to fully utilize pre-trained models in AI-based drug development to maximize their migration to downstream tasks with small sample sizes and improve model performance.

[0058] The embodiment of the present application makes full use of the existing pre-trained model for molecules, so that it can complete fine-tuning on small sample size tasks in the field of drug development to the greatest extent possible, thereby improving model performance. Specifically, first, the molecular data that meets the similarity requirements in the application domain is searched using a similarity metric. Then, a small-scale pre-training is performed using the molecular data that meets the similarity requirements in the application domain. Next, fine-tuning is performed on downstream tasks for actual supervision tasks. The above method effectively improves the prediction accuracy of the pre-trained model migrated to the target field.

[0059] The following will be passed Figures 1 to 19 A data processing method, device, model training method, device and electronic device in an embodiment of the present application are described in detail.

[0060] Figure 1 The following schematically illustrates an exemplary system architecture to which the data processing method, device, model training method, device, and electronic device according to the embodiments of the present application can be applied. It should be noted that: Figure 1 What is shown is merely an example of a system architecture to which the embodiments of the present application can be applied, in order to help those skilled in the art understand the technical content of the present application, but it does not mean that the embodiments of the present application cannot be used in other devices, systems, environments or scenarios.

[0061] See also Figure 1 According to the embodiment, the system architecture 100 may include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0062] The user can use the terminal devices 101, 102, 103 to interact with other terminal devices and the server 105 through the network 104 to receive or send information, such as sending a target data acquisition request, a model training request, a molecular property prediction request, a molecular design request, and receiving target data, model training results, molecular properties, and molecular structures, etc. The terminal devices 101, 102, 103 can be installed with various communication client applications, such as drug development applications, material design applications, web browser applications, database applications, search applications, instant messaging tools, email clients, social platform software, and other applications.

[0063] The terminal devices 101, 102, 103 include but are not limited to smart desktop computers, tablet computers, laptop computers, and other electronic devices that can support functions such as surfing the Internet, modeling, analytical calculation, and design.

[0064] The server 105 can receive a target data acquisition request, a model training request, a molecular property prediction request, a molecular design request, etc., and can also be used to send target data, model training results, molecular properties and molecular structures, etc. to the terminal devices 101, 102, 103. For example, the server 105 can be a background management server, a server cluster, etc.

[0065] It should be noted that the number of terminal devices, networks and servers is only illustrative, and any number of terminal devices, networks and clouds may be provided according to implementation requirements.

[0066] Figure 2 The flowchart of the data processing method according to the embodiment of the present application is schematically shown.

[0067] like Figure 2 As shown, this embodiment provides a method for data processing, which includes operations S210 to S220, which are specifically as follows.

[0068] In operation S210, first features of m training data in a training data set are obtained, where m is an integer greater than or equal to 1.

[0069] The training data in the training data set may have label information or no label information. The training data may include molecular data, and accordingly, the training data set may include a molecular data set. For example, the training data set may be used to fine-tune a pre-trained model. For another example, the training data set may be used for pre-training.

[0070] For example, an unlabeled training data set can be used to perform self-supervised learning training on a neural network. An unlabeled training data set can be obtained through a variety of channels. For example, it can be a collection of all training data sets obtained free of charge through public channels, or it can be a collection of small molecule compound data obtained through commercial purchase. For example, an unlabeled training data set can be a commercially purchased ZINC15 data set (containing approximately 20 million small molecule compound data). For example, an unlabeled training data set can be a molecular data set accumulated by a user based on historical research and development experience. For example, it can also be a collection of the above two sets.

[0071] It should be noted that the training data set may also include drug-like small molecule data, which can also be obtained through a variety of channels. For example, it can be a collection of all drug-like small molecule data obtained free of charge through public channels, such as the CHEMBL data set of drug-like small molecule data with biological activity.

[0072] The first feature of the training data may be a feature automatically learned based on deep learning technology, for example, a feature extracted by a neural network. For example, the method of extracting features based on a pre-trained model includes, but is not limited to: extracting the last layer hidden variables of the pre-trained model. In addition, the last layer hidden variables may be globally pooled, wherein the global pooling is one of the following: global average pooling or global maximum pooling. In addition, the first feature may also be a preset feature based on expert experience, such as a feature that can be used to characterize molecular properties. The first feature may also be a feature obtained by combining the above two methods.

[0073] In operation S220, n target data having the second feature are obtained so as to adjust the model parameters of the pre-trained model based on at least part of the n target data, where n is an integer greater than or equal to 1.

[0074] Among them, the first feature and the second feature meet the similarity requirement, the pre-trained model is trained based on the above training data set, and the target data has no label information or has label information. It should be noted that this technical solution is particularly suitable for target data without label information, or target data without corresponding label information.

[0075] In some embodiments, the first feature includes a first fingerprint feature, and the second feature includes a second fingerprint feature. The similarity requirement may include: the similarity between the first fingerprint feature and the second fingerprint feature is greater than or equal to a preset similarity threshold. It should be noted that as long as it can indicate / represent a molecule (fingerprint / descriptor), etc., it can be used as the first feature or the second feature, and no limitation is made here. Among them, the preset similarity threshold can be determined based on expert experience or usage effect, such as 50%, 60%, 70%, 80%, 90%, 95%, 98%, etc.

[0076] For example, the first feature and the second feature may characterize whether certain specific structures exist in the molecular structure corresponding to the molecular data.

[0077] Figure 3 The figure schematically shows a first fingerprint feature according to an embodiment of the present application.

[0078] See also Figure 3The first fingerprint feature can be a string, such as a binary data string. Each bit corresponds to a specific structure, such as bit 1 corresponds to a benzene ring structure, bit 2 corresponds to a HC≡N structure, bit 3 corresponds to a COCl2 structure, bit 4 corresponds to a CH2O structure, bit 5 corresponds to a N2H4 structure, and bit 6 corresponds to NH4 + Structure, etc. are not listed here. It should be noted that the total number of bits, structure, and the correspondence between each bit and the molecular structure are only exemplary and cannot be understood as a limitation of the present disclosure. For example, if the value of bit 1 is 1, it indicates that the molecule corresponding to the molecular data includes a benzene ring structure. For example, if the value of bit 2 is 0, it indicates that the molecule corresponding to the molecular data does not include a HC≡N structure.

[0079] It should be noted that the above fingerprint features are described in the form of binary data, which should not be understood as a limitation of the present application. As long as it can indicate / represent a molecule (fingerprint / descriptor), it can be used. Morgan fingerprint (corresponding to binary) is just one of the optional methods.

[0080] Figure 4 The schematic diagram of the molecular structure according to the embodiment of the present application is schematically shown.

[0081] See also Figure 4 , showing two molecular structures. Among them, the molecular structure can be represented by the simplified molecular input line entry system (SMILES). SMILES is a specification that explicitly describes the molecular structure using ASCII strings. For example, the molecular structure on the left can be represented as ClC1CNCOC1. The molecular structure on the right can be represented as BrC1CNCOC1. The features of the molecular structure corresponding to the molecular data can be extracted through the above ASCII strings, such as Figure 4 Both molecular structures include the structure: C1CNCOC1, so as to determine the similarity of the molecular structures, and then facilitate the determination of the similarity between the molecular data.

[0082] It should be noted that, in addition to expressing the molecular structure as SMILES, the molecule can also be expressed as a low-dimensional descriptor such as MACCS, Morgan, and hybrid fingerprint, which is not limited here.

[0083] This embodiment obtains target data with high similarity to the training data in the training set from a large amount of unlabeled data. This allows the pre-trained model trained with the training data set to be fine-tuned once in an unsupervised manner based on the target data, and then the obtained pre-trained model is fine-tuned on the downstream supervision task. This embodiment can effectively improve the model training effect, such as improving the prediction accuracy for downstream tasks.

[0084] The following is an exemplary description of the process of acquiring target data.

[0085] In some embodiments, the first fingerprint feature includes first binary data, the second fingerprint feature includes second binary data, and each bit in the first binary data and the second binary data represents a substructure of a molecule. For example, at least one structural feature in the current molecule data can be extracted from the above ASCII string, and the structure can be expressed as follows: Figure 3 The data format of the first feature shown encodes the extracted at least one structural feature to obtain a second feature, such as a binary data string. In this way, the similarity between two molecular data can be determined by calculating the similarity between the two binary data strings.

[0086] Accordingly, obtaining n target data with the second feature may include the following operations. First, a candidate data set is obtained, wherein the similarity between the second binary data of each candidate data in the candidate data set and the first binary data is greater than or equal to a preset similarity threshold, and the application domain of the candidate data set is the same as the application domain of the training data set. Then, n candidate data are obtained from the candidate data set.

[0087] The molecules corresponding to the molecular data can be applied to at least one application domain. Among them, the application domain can refer to the characteristic distribution that conforms to the molecular structure in the labeled data set. For example, the application domain can refer to the field where the molecule can be applied, such as drug treatment, imaging, health care, etc. Of course, the application domain can also be subdivided. For example, the application domain of drug treatment can be further divided into: infection treatment, specific disease treatment, immunity enhancement treatment, etc. Among them, infection treatment can be further divided into: viral infection treatment, bacterial infection treatment, and specific types of viral infection treatment. Specific disease treatment can be further divided into: antibiotics, cardiovascular and cerebrovascular drugs, digestive system drugs, respiratory system drugs, urinary system drugs, blood system drugs, etc. In summary, the application domain corresponding to the training data set can be determined.

[0088] By selecting target data from the same application domain as the training data set, it helps to improve the robustness of the model. By pre-constructing multiple candidate data sets, each data set has its own characteristics, it is convenient to determine the target data set from multiple candidate data sets based on the first characteristic, such as the second characteristic of the target data set has a high similarity with the first characteristic.

[0089] Figure 5 A schematic diagram of a process for obtaining target training data according to an embodiment of the present application is schematically shown.

[0090] See also Figure 5 , taking application domain 1 as an example, multiple training data in application domain 1 can be classified to obtain multiple sub-data sets, as shown in the dashed boxes. The training data in each dashed box has certain common features, such as having a benzene ring. These common features can be encoded as binary data for characterization. For example, Figure 5 The sub-datasets where training data 1, 2, 7, and 8 are located can be characterized as having the characteristics of the first binary data. The sub-datasets where training data 4, 5, 6, 10, 11, and 12 are located can be characterized as having the characteristics of the second binary data. The sub-datasets where training data 8, 9, 14, and 15 are located can be characterized as having the characteristics of the third binary data.

[0091] It should be noted that the above-mentioned multiple sub-datasets are stored in the same or adjacent storage space. The above-mentioned multiple sub-datasets can also be stored in different storage spaces (such as different databases, data tables), which is not limited here.

[0092] This embodiment effectively improves the convenience of obtaining target data by selecting a candidate data set and selecting a specific number of candidate data from the candidate data set, and facilitates the use of the target data (set) for model training.

[0093] In some embodiments, the first fingerprint feature includes first binary data, the second fingerprint feature includes second binary data, and each bit in the first binary data and the second binary data represents a substructure of a molecule.

[0094] Accordingly, for each target data, obtaining n target data with a second feature may include the following operations. First, candidate data is obtained, and the application domain of the candidate data is the same as the application domain of the training data set. Then, the second binary data of the second fingerprint feature of the candidate data is obtained, and then, if the similarity between the first binary data and the second binary data is greater than or equal to a preset similarity threshold, the candidate data is used as the target data. This can improve the flexibility of obtaining target data, such as randomly or sequentially obtaining a certain candidate data, and then calculating the similarity between the candidate data and the training data one by one until the user needs are met.

[0095] In some embodiments, if the similarity between the first binary data and the second binary data is greater than or equal to a preset similarity threshold, taking the candidate data as the target data may include the following operations.

[0096] If the ratio between the first total number of digits and the second total number of digits is greater than or equal to a preset similarity threshold, the candidate data is taken as the target data.

[0097] For example, the first total number of bits is the total number of bits where the same bit is 1 in the first binary data and the second binary data, and the second total number of bits is the difference between the total number of bits of 1 in the first binary data and the second binary data and the first total number of bits.

[0098] For another example, the first total number of bits is the total number of bits where the same bit is 0 in the first binary data and the second binary data, and the second total number of bits is the difference between the total number of bits of 0 in the first binary data and the second binary data and the first total number of bits.

[0099] Figure 6 A schematic diagram of another process of obtaining target training data according to an embodiment of the present application is schematically shown.

[0100] See also Figure 6 , the user can first select training data 1 from the training data for application domain 2, use it as candidate data 1, and then calculate the similarity between the second feature and the first feature of candidate data 1. If the similarity does not meet the similarity requirement, it can be determined that training data 1 is non-target data. For example, the user can also select training data 5 as candidate data 5, and then calculate the similarity between the second feature and the first feature of candidate data 5. If the similarity meets the similarity requirement, it can be determined that training data 1 is target data.

[0101] In a specific embodiment, a similarity search based on Morgan Fingerprints can be used to search in an internally prepared commercial database and the training set similarity is in the Tanimoto similarity>=preset similarity threshold, such as 0.5, 0.6, 0.65, 0.7, 0.8, 0.9, etc. Morgan fingerprint is a binary fingerprint, each bit represents a substructure, if the fingerprint is 1, it means that the molecule has this substructure. Tanimoto similarity is achieved by comparing whether each bit of the binary fingerprint is the same. For example, the same number of fingerprint bits is 1 / (the sum of the number of fingerprint bits that are 1 for both - the same number of fingerprint bits is 1), as shown in formula (1).

[0102]

[0103] Among them, a is the Morgan fingerprint of molecule 1, and b is the Morgan fingerprint of molecule 2.

[0104] Morgan fingerprint is a circular fingerprint, also a topological fingerprint, obtained by modifying the standard Morgan algorithm. It can be roughly equivalent to the Extended-Connectivity Fingerprints (ECFPs). The use of Morgan fingerprints can effectively improve the calculation speed. In addition, it does not need to be predefined (it can represent an infinite number of different molecular features), can contain chiral information, each element in the fingerprint represents a specific substructure, can be easily analyzed and interpreted, and can be modified accordingly according to different needs. In addition, it can also be used for similarity search, clustering, virtual screening and other directions.

[0105] Morgan fingerprints can be generated in the following way. First, atoms are initialized. Each heavy atom is assigned an integer identifier. Then, iterative updates are performed, such as merging the heavy atoms around each heavy atom until the specified radius is reached. Next, features are generated, such as operating on substructures and generating a feature list.

[0106] In addition, the target data may also be determined in other ways. In some embodiments, the training data includes molecular structure data.

[0107] Accordingly, obtaining the n target data having the second characteristic may include: obtaining the n target data having the second characteristic by searching for the largest common molecular substructure.

[0108] Figure 7 The following is a schematic diagram of a process for determining a maximum common substructure according to an embodiment of the present application.

[0109] See also Figure 7 , both molecular structure diagrams above include the common substructure shown below.

[0110] For example, obtaining n target data with the second feature by searching for the largest common molecular substructure may include the following operations.

[0111] First, a candidate molecule dataset is obtained, and the application domain of the candidate molecule dataset is the same as the application domain of the training data.

[0112] Then, p molecular graphs corresponding to at least part of the molecular data in the candidate molecular data set are determined, where p is a positive integer greater than or equal to 0, and p is greater than or equal to n.

[0113] Next, the isomorphic parts of at least some of the p molecular graphs and the number of atoms in the isomorphic parts are determined.

[0114] Then, based on the similarity between the number of atoms of the candidate molecule in the candidate molecule dataset and the number of atoms of the isomorphic part, n target data are determined from the candidate molecule dataset.

[0115] For example, rdkit provides a function to find the maximum common substructure (MCS) between multiple molecules. This function can be used in scenarios such as searching for the same skeleton and finding pharmacophores.

[0116] In the adaptive fine-tuning of this embodiment, the target data can be obtained by the maximum common substructure search method, by converting (at least part of the application domain) molecules into molecular graphs respectively, and then retrieving the most isomorphic parts in these molecular graphs and calculating their atomic numbers, and then substituting the atomic number and the common atomic number of each molecule in the molecule into the Tanimoto similarity to calculate the similarity.

[0117] In addition, the similarity between molecular data may also be calculated based on Euclidean distance, etc., which is not limited here.

[0118] Another aspect of the present application also provides a model training method.

[0119] Figure 8 The flowchart of the model training method according to an embodiment of the present application is schematically shown.

[0120] See also Figure 8 , the model training method may include operations S810 to S820.

[0121] In operation S810 , a first target data set is obtained, where the first target data set includes target data obtained by the above method, and the target data has no label information.

[0122] In operation S820, the pre-trained model is trained using the first target data set to adjust model parameters of the pre-trained model to obtain an adjusted pre-trained model.

[0123] In some embodiments, in the process of training the pre-trained model using the first target data set, at least one training task of context prediction and attribute masking can be added. In this embodiment, in the process of self-supervised learning training of the pre-trained model using the unlabeled first target data set, at least the following training tasks can be added: context prediction, attribute masking, and graph-level motif prediction.

[0124] For example, attribute masking is performed by masking the attributes of atoms or edges and predicting the masked content at that position. Specifically, 10% of the identifiers (tokens) in each input string can be randomly masked. In order to prevent overfitting, 5 training process epochs (forward calculation + back propagation) can be trained in an early stopping manner. When learning to recover the masked identifier (masktoken), the model forms a representation topology of the chemical space, which can be applied to different molecular attribute prediction tasks.

[0125] For example, context prediction is used to predict the context words corresponding to a given center word. For example, in Skip-gram, the center word is the input word and the context words are the output words.

[0126] In this embodiment, the pre-trained model can be the following two models: a model obtained by assigning parameters of an existing model (eg, an open source model), or a pre-trained model based on a migrated open source model.

[0127] For example, an initial model is pre-trained on an original task, and then the model is used on the target task. According to the characteristics of the target task, target data similar to the characteristics of the training data required for the target task are selected, so that the initial model can be pre-trained again based on the target data, which helps to improve the purpose of achieving the target task.

[0128] The architecture of the above-mentioned pre-trained models includes but is not limited to: Graph Isomorphism Network (GIN), Graph Attention Network (GAT), Graph Aggregation Network (GraphSage), and Graph Convolutional Network (GCN).

[0129] For example, graph representation learning of graph isomorphic networks mainly includes the following two processes: first, the node representation is calculated. Then, the representation of each node on the graph is pooled (graph pooling), or called graph readout, to obtain the graph representation.

[0130] In a specific embodiment, the first target data set (unlabeled data) is obtained based on the pre-training task of atom masking to obtain a model that has been transferred on the pre-training task. During the training process, the properties of atoms or edges are masked, and the content of the masked position is predicted to achieve a secondary pre-training process.

[0131] For ease of understanding, the first pre-training process is described as follows. The original model can be a Contextpred model or a GROVER model, etc. For the Contextpred model, two self-supervised tasks including context prediction and attribute masking can be added. For example, predict the adjacent nodes within a certain range around the central node and predict the randomly masked nodes / edges. For the GROVER model, two unsupervised tasks including context prediction and graph-level motif (important functional fragments in small molecules) prediction can be added. For example, input the target atom / bond to predict the adjacent atom / bond information and predict whether the motif detected by Rdkit (Rational Discovery Kit) appears. Through pre-training self-supervised learning, the Contextpred model and the GROVER model have a grasp of the global characteristics of the molecular space, realizing the first pre-training process.

[0132] In addition, in practical applications, the model fine-tuned through the secondary pre-training process can also be used for the migration of downstream supervised data.

[0133] In some embodiments, the method may further include the following operation: training the adjusted pre-trained model using training data of a specific application domain to adjust model parameters of the adjusted pre-trained model, wherein the training data of the specific application domain has label information, and the specific application domain is the same as or different from the application domain of the training data set. When the specific application domain is the same as the application domain of the training data set, the training effect is better.

[0134] Fig. 9 The schematic diagram shows an architecture diagram of a model training system according to an embodiment of the present application.

[0135] See also Fig. 9, data in the application domain with high similarity to the training data can be found from the commercial library as the target data, wherein the training data can be data used to fine-tune the pre-trained model. Before fine-tuning the pre-trained model, first use a small amount of target data to perform unsupervised learning on the pre-trained model, such as atomic masking, to obtain a model after adaptive domain migration. Then, the model after adaptive domain migration is fine-tuned using the above training data to obtain a fine-tuned downstream model. The training data can have label information or no label information.

[0136] In some embodiments, training the adjusted pre-trained models using training data of a specific application domain may include: adding a linear classifier or a multilayer perceptron to the adjusted multiple pre-trained models to obtain one or more molecular prediction models or molecular evaluation models to be adjusted, wherein the adjusted multiple pre-trained models are trained according to the above model training method.

[0137] It should be noted that different functions such as molecular property prediction or molecular evaluation can be achieved for different training data and model structures.

[0138] For example, a linear classifier or a multilayer perceptron is added to one or more molecular pre-trained models to obtain one or more pre-trained models to be adjusted, and then the parameters of the one or more pre-trained models to be adjusted are adjusted using labeled molecular data sets to obtain one or more adjusted molecular models.

[0139] For example, the adjusted pre-trained model can be used for a molecular prediction task. Specifically, the model parameters of one or more molecular prediction models to be adjusted can be adjusted using target data without label information to obtain one or more adjusted molecular prediction models.

[0140] For example, the adjusted pre-trained model can be used for the molecular evaluation task. Specifically, the model parameters of one or more molecular evaluation models to be adjusted can be adjusted using target data with label information to obtain one or more adjusted molecular evaluation models.

[0141] Fig.10 A schematic diagram of a model training process according to an embodiment of the present application is schematically shown.

[0142] See also Fig.10, use unlabeled training data for pre-training to obtain a pre-trained model. Use the target data of the application domain (such as unlabeled data) to perform unsupervised learning on the pre-trained model again to obtain a transferred model. Then, the transferred model can be fine-tuned using the unlabeled training data of the application domain to obtain an adjusted model, which can be used for prediction tasks, etc. In addition, the transferred model can be fine-tuned using labeled training data of the application domain to obtain an adjusted model, which can be used for evaluation tasks, etc.

[0143] In this embodiment, the target data is the training data in the application domain, which can achieve the effect of a model similar to that of secondary pre-training. Compared with the pre-training in the previous pre-training framework, it can have the following differences: the primary data set is large (massive (tens of millions) unlabeled data in the database), and the secondary data set is small, but targeted at a specific application domain. Therefore, the model training effect can be effectively improved without adding too many training tasks.

[0144] In order to determine the effect of the adjusted pre-trained model obtained in each embodiment, the test is as follows.

[0145] Specifically, the datasets used include datasets of physicochemical properties from MoleculeNet and the applicant's internal data sets, as shown in Table 1. Table 1 includes the Freesolv dataset that records the solvation energy of molecules, the ESOL dataset that records the solubility of molecules, the Caco2 dataset that records the permeability of molecules, and the Lipolicity dataset that records the lipid solubility of molecules.

[0146] Table 1

[0147]

[0148]

[0149] Among them, the FreeSolv dataset contains experimental values ​​of the hydration free energy of small molecules in water, which can be obtained through molecular dynamics simulation. The ESOL dataset contains water solubility data of common organic small molecules (for example: logarithmic solubility, mol / L). The Lipophilicity dataset contains experimental results of the octanol or water partition coefficients of known molecules, which reflect the solubility of the molecules. In addition, the datasets shown below can also be used. For example, the BBBP dataset contains a dataset of molecules that have been measured to penetrate the blood-brain barrier. The ClinTox dataset includes data on drugs approved by the FDA and those that have failed clinical trials due to toxicity. The Tox21 dataset includes data on the toxicity of compounds that have been measured on 12 different targets, including experimental measurements of nuclear receptors and stress response pathways. The HIV dataset includes data on small molecules that have been experimentally determined to have the ability to inhibit HIV replication.

[0150] According to the technical route mentioned above, experiments were conducted on the above datasets, and the test verification results are shown in Tables 2 to 4. The above datasets are all datasets for regression tasks, and RMSE and R2 are used as evaluation indicators for the tasks. The improvements on Freesolv and Lipolicity are relatively high. For example, the RMSE is 3.589% and 0.803% respectively, the ESOL dataset has dropped by 0.574%, and Caco2 has maintained the same effect. From the results, adaptive fine-tuning (using the target data to train the pre-training data and then perform the fine-tuning task) is more effective in improving, at least it can keep the performance on par with the original model.

[0151] The above-mentioned preset evaluation indicators may include at least one of the following: evaluation indicator determination coefficient, area under the ROC curve, accuracy, recall rate, sensitivity, specificity, root mean square error, and mean absolute error.

[0152] The calculation formula of the root mean square error RMSE is shown in formula (2).

[0153]

[0154] Where m is the number of samples, y i is the true value, is the predicted value;

[0155] Coefficient of determination R 2 The calculation formula of is shown in formula (3).

[0156] R 2 =SSR / SST=1-SSE / SST Formula (3)

[0157] Among them, the total sum of squares (SST), the regression sum of squares (SSR), and the error sum of squares (SSE).

[0158] It should be noted that the mean value (MEAN) and root mean square error (STD) can also be used for evaluation.

[0159] Table 2

[0160] Freesolv MEAN STD Improvement ratio (%) RMSE 1.436 0.028 3.856 R2 0.905 0.004 0.959

[0161] Table 3

[0162] Caco2 MEAN STD Improvement ratio (%) RMSE 0.530 0.004 -0.003 R2 0.550 0.005 -0.013

[0163] Table 4

[0164] Lipolicity MEAN STD Improvement ratio (%) RMSE 0.544 0.007 0.803 R2 0.785 0.005 0.358

[0165] In general, when the task is improved, the indicators improved by adaptive fine-tuning are relatively high (see Table 2 and Table 4), and the effect does not decrease significantly when it is not improved (see Table 3), indicating that adaptive fine-tuning has non-negative auxiliary effect on fine-tuning the pre-trained model to downstream tasks.

[0166] Another aspect of the present application provides a method for predicting molecular properties.

[0167] Fig.11 The flowchart of the molecular property prediction method according to an embodiment of the present application is schematically shown.

[0168] See also Fig.11 , the molecular property prediction method may include operations S1110 to S1120.

[0169] In operation S1110 , molecule data is obtained.

[0170] In operation S1120, based on the molecular data, a molecular prediction model obtained by using the adjusted pre-trained model as described above is used to predict molecular properties of the molecular data.

[0171] The molecular attribute can be a structure-function score of the molecule based on data. In this application, the predicted molecular attributes include, but are not limited to, water-soluble molecular physiological, biochemical, and biophysical properties. When predicting molecular attributes, a fine-tuned model is required. Then, the molecular data is input into the fine-tuned model to obtain the molecular attribute prediction.

[0172] In some embodiments, the above method may further include: evaluating the molecular data using a molecular evaluation model obtained by the adjusted pre-training model to obtain a molecular evaluation result of the molecular data.

[0173] For example, whether the molecule corresponding to the current molecular data is a potential effective drug for a certain disease. The evaluation result can be yes or no, or a scoring result, which is not limited here.

[0174] Another aspect of the present application provides a method for predicting molecular properties.

[0175] Fig.12 The flowchart of the molecular property evaluation method according to an embodiment of the present application is schematically shown.

[0176] See also Fig.12 , the molecular property prediction method may include operations S1210 to S1220.

[0177] In operation S1210, molecule data is obtained, the molecule data having tag information.

[0178] In operation S1220, based on the molecular data, the molecular evaluation model obtained by the adjusted pre-training model as described above is used for evaluation to obtain a molecular evaluation result of the molecular data. The molecular evaluation result may be as shown above.

[0179] This embodiment uses similar unlabeled internal molecular structure data to first fine-tune the upstream pre-trained model in an unsupervised manner and then fine-tune the downstream tasks, effectively improving the accuracy of molecular property prediction and molecular evaluation.

[0180] On the other hand, the present application also provides a design method.

[0181] Fig.13 The flowchart of the design method according to the embodiment of the present application is schematically shown.

[0182] See also Fig.13 , the above-mentioned design method may include operations S1310 to S1320.

[0183] In operation S1310, molecular attributes and / or molecular evaluation results are obtained. For details, please refer to the above related embodiments, which will not be described in detail here.

[0184] In operation S1320 , drug design or material design is performed based on the molecular attributes and / or the molecular evaluation results.

[0185] This embodiment uses molecular properties and molecular evaluation results with high confidence levels for design, such as pharmaceutical development, which can effectively improve design efficiency and success rate.

[0186] On the other hand, the present application also provides a data processing device.

[0187] Fig.14 The block diagram schematically shows a data processing device according to an embodiment of the present application.

[0188] See also Fig.14 , the data processing device 1400 may include: a first feature acquisition module 1410 and a target data acquisition module 1420.

[0189] The first feature obtaining module 1410 is used to obtain the first feature of m training data in the training data set, where m is an integer greater than or equal to 1.

[0190] The target data acquisition module 1420 is used to obtain n target data with a second feature, so as to adjust the model parameters of the pre-trained model based on at least part of the n target data, where n is an integer greater than or equal to 1. The first feature and the second feature meet the similarity requirement, the pre-trained model is trained based on the training data set, and the target data has no label information or has label information.

[0191] In some embodiments, the first feature includes a first fingerprint feature and the second feature includes a second fingerprint feature.

[0192] The similarity requirement includes: the similarity between the first fingerprint feature and the second fingerprint feature is greater than or equal to a preset similarity threshold.

[0193] In some embodiments, the first fingerprint feature includes first binary data, the second fingerprint feature includes second binary data, and each bit in the first binary data and the second binary data represents a substructure of a molecule.

[0194] Correspondingly, the target data obtaining module 1420 includes: a first candidate data set obtaining unit and a candidate data selecting unit.

[0195] The first candidate data set obtaining unit is used to obtain a candidate data set, wherein the similarity between the second binary data of each candidate data in the candidate data set and the first binary data is greater than or equal to a preset similarity threshold, and the application domain of the candidate data set is the same as the application domain of the training data set.

[0196] The candidate data selection unit is used to obtain n candidate data from the candidate data set.

[0197] In some embodiments, the first fingerprint feature includes first binary data, the second fingerprint feature includes second binary data, and each bit in the first binary data and the second binary data represents a substructure of a molecule.

[0198] Accordingly, the target data obtaining module 1420 includes: a second candidate data set obtaining unit, a binary data obtaining unit and a comparing unit.

[0199] The second candidate data set obtaining unit is used to obtain candidate data, and the application domain of the candidate data is the same as the application domain of the training data set.

[0200] The binary data obtaining unit is used to obtain second binary data of the second fingerprint feature of the candidate data.

[0201] The comparison unit is configured to use the candidate data as the target data if the similarity between the first binary data and the second binary data is greater than or equal to a preset similarity threshold.

[0202] In some embodiments, the comparison unit has a function for taking the candidate data as the target data if the ratio between the first total number of digits and the second total number of digits is greater than or equal to a preset similarity threshold. The first total number of digits is the total number of digits where the same bit is 1 in the first binary data and the second binary data, and the second total number of digits is the difference between the total number of digits of 1 in the first binary data and the second binary data and the first total number of digits. Alternatively, the first total number of digits is the total number of digits where the same bit is 0 in the first binary data and the second binary data, and the second total number of digits is the difference between the total number of digits of 0 in the first binary data and the second binary data and the first total number of digits.

[0203] In certain embodiments, the training data includes molecular structure data.

[0204] Accordingly, the target data acquisition module 1420 includes: a common molecule search unit, configured to obtain n target data having the second feature by searching for the largest common molecule substructure.

[0205] In some embodiments, the public molecule search unit has a function of: first, obtaining a candidate molecule data set, the application domain of the candidate molecule data set being the same as the application domain of the training data. Then, determining p molecular graphs corresponding to at least part of the molecular data in the candidate molecule data set, where p is a positive integer greater than or equal to 0, and p is greater than or equal to n. Next, determining the isomorphic part of at least part of the molecular graphs in the p molecular graphs and the number of atoms of the isomorphic part. Then, based on the similarity between the number of atoms of the candidate molecule in the candidate molecule data set and the number of atoms of the isomorphic part, determining n target data from the candidate molecule data set.

[0206] On the other hand, the present application also provides a training device.

[0207] Fig.15 A block diagram of a training device according to an embodiment of the present application is schematically shown.

[0208] See also Fig.15 , the training device 1500 may include: a data set acquisition module 1510 and a model training module 1520.

[0209] The data set acquisition module 1510 is used to obtain a first target data set, where the first target data set includes target data obtained according to the above apparatus 1400, and the target data has no label information.

[0210] The model training module 1520 is used to train the pre-trained model using the first target data set to adjust the model parameters of the pre-trained model to obtain an adjusted pre-trained model.

[0211] In some embodiments, the above-mentioned training device 1500 may also include: a model parameter adjustment module, which is used to train the adjusted pre-trained model using training data of a specific application domain to adjust the model parameters of the adjusted pre-trained model, wherein the training data of the specific application domain has label information, and the specific application domain is the same as or different from the application domain of the training data set.

[0212] In certain embodiments, the above-mentioned model parameter adjustment module includes: a model structure adjustment unit, which is used to add a linear classifier or a multilayer perceptron to the adjusted multiple pre-trained models to obtain one or more molecular prediction models or molecular evaluation models to be adjusted, wherein the adjusted multiple pre-trained models are trained according to the above-mentioned model training method.

[0213] The target data without label information is used to adjust the model parameters of one or more molecular prediction models to be adjusted, so as to obtain one or more adjusted molecular prediction models.

[0214] Alternatively, the target data with label information is used to adjust the model parameters of one or more molecular evaluation models to be adjusted to obtain one or more adjusted molecular evaluation models.

[0215] In some embodiments, the training task includes: adding at least one of context prediction and attribute masking.

[0216] On the other hand, the present application also provides a prediction device.

[0217] Fig.16 The block diagram of the prediction device according to the embodiment of the present application is schematically shown.

[0218] See also Fig.16 , the prediction device 1600 may include: a first molecule data acquisition module 1610 and a molecule property prediction module 1620 .

[0219] The first molecule data obtaining module 1610 is used to obtain molecule data.

[0220] The molecular property prediction module 1620 is used to predict the molecular properties of the molecular data based on the molecular data by processing the molecular data using the molecular prediction model adjusted as described above.

[0221] In some embodiments, the prediction device 1600 may further include: an evaluation model training module for evaluating the molecular data based on the molecular data using a molecular evaluation model obtained by the adjusted pre-training model as described above to obtain a molecular evaluation result of the molecular data.

[0222] On the other hand, the present application also provides an evaluation device.

[0223] Fig.17 A block diagram of an evaluation device according to an embodiment of the present application is schematically shown.

[0224] See also Fig.17 The evaluation device 1700 may include: a second molecule data obtaining module 1710 and a molecule property evaluation module 1720 .

[0225] The second molecular data obtaining module 1710 is used to obtain molecular data, and the molecular data has label information.

[0226] The molecular property evaluation module 1720 is used to evaluate the molecular data using the molecular evaluation model obtained by the adjusted pre-training model as described above, to obtain the molecular evaluation result of the molecular data.

[0227] On the other hand, the present application also provides a design device.

[0228] Fig.18 A block diagram of a design device according to an embodiment of the present application is schematically shown.

[0229] See also Fig.18 , the prediction device 1800 may include: an information acquisition module 1810 and a design module 1820.

[0230] The information acquisition module 1810 is used to obtain molecular properties according to the above molecular property prediction device, and / or to obtain molecular evaluation results according to the above molecular property evaluation device.

[0231] The design module 1820 is used to perform drug design or material design based on molecular properties and / or molecular evaluation results.

[0232] Regarding the devices 1400, 1500, 1600, 1700, and 1800 in the above embodiments, the specific manner in which each module and unit performs operations has been described in detail in the embodiments of the method, and will not be elaborated again here.

[0233] Another aspect of the present application provides an electronic device.

[0234] Fig.19 A block diagram of an electronic device according to an embodiment of the present application is schematically shown.

[0235] See also Fig.19 , the electronic device 1900 includes a memory 1910 and a processor 1920 .

[0236] The processor 1920 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

[0237] The memory 1910 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. Among them, ROM can store static data or instructions required by the processor 1920 or other modules of the computer. The permanent storage device may be a readable and writable storage device. The permanent storage device may be a non-volatile storage device that does not lose the stored instructions and data even after the computer is powered off. In some embodiments, the permanent storage device uses a large-capacity storage device (such as a magnetic or optical disk, flash memory) as a permanent storage device. In some other embodiments, the permanent storage device may be a removable storage device (such as a floppy disk, optical drive). The system memory may be a readable and writable storage device or a volatile readable and writable storage device, such as a dynamic random access memory. The system memory may store some or all instructions and data required by the processor at run time. In addition, the memory 1910 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (such as DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks may also be used. In some embodiments, the memory 1910 may include a readable and / or writable removable storage device, such as a laser disc (CD), a read-only digital versatile disc (such as a DVD-ROM, a double-layer DVD-ROM), a read-only Blu-ray disc, an ultra-density optical disc, a flash memory card (such as an SD card, a mini SD card, a Micro-SD card, etc.), a magnetic floppy disk, etc. The computer-readable storage medium does not include carrier waves and transient electronic signals transmitted wirelessly or wired.

[0238] The memory 1910 stores executable codes, and when the executable codes are processed by the processor 1920 , the processor 1920 can execute part or all of the methods described above.

[0239] In addition, the method according to the present application may also be implemented as a computer program or a computer program product, which includes computer program code instructions for executing some or all of the steps in the above method of the present application.

[0240] Alternatively, the present application can also be implemented as a computer-readable storage medium (or non-transitory machine-readable storage medium or machine-readable storage medium) on which executable code (or computer program or computer instruction code) is stored. When the executable code (or computer program or computer instruction code) is executed by a processor of an electronic device (or server, etc.), the processor executes part or all of the steps of the above-mentioned method according to the present application.

[0241] The embodiments of the present application have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A data processing method, characterized in that: include: Obtain the first feature of m training data in the training data set, where m is an integer greater than or equal to 1; Obtaining n target data having a second feature, so as to adjust a model parameter of a pre-trained model based on at least part of the n target data, where n is an integer greater than or equal to 1; Wherein, the first feature and the second feature meet the similarity requirement, the pre-trained model is trained based on the training data set, and the target data has no label information or has label information; The first feature includes a first fingerprint feature, and the second feature includes a second fingerprint feature; The similarity requirement includes: the similarity between the first fingerprint feature and the second fingerprint feature is greater than or equal to a preset similarity threshold; The first fingerprint feature includes first binary data, the second fingerprint feature includes second binary data, and each bit in the first binary data and the second binary data represents a substructure of a molecule; The obtaining of n target data having the second feature comprises: Obtaining a candidate data set, wherein the similarity between the second binary data of each candidate data in the candidate data set and the first binary data is greater than or equal to the preset similarity threshold, and the application domain of the candidate data set is the same as the application domain of the training data set; Obtain n candidate data from the candidate data set; or The first fingerprint feature includes first binary data, the second fingerprint feature includes second binary data, and each bit in the first binary data and the second binary data represents a substructure of a molecule; The obtaining of n target data having the second feature comprises: for each target data, Obtaining candidate data, wherein the application domain of the candidate data is the same as the application domain of the training data set; Obtain second binary data of a second fingerprint feature of the candidate data; If the similarity between the first binary data and the second binary data is greater than or equal to the preset similarity threshold, the candidate data is used as the target data.

2. The data processing method according to claim 1, characterized in that: If the similarity between the first binary data and the second binary data is greater than or equal to the preset similarity threshold, taking the candidate data as the target data includes: If the ratio between the first total number of digits and the second total number of digits is greater than or equal to the preset similarity threshold, taking the candidate data as the target data; The first total number of digits is the total number of digits where the same bit in the first binary data and the second binary data is 1, and the second total number of digits is the difference between the total number of digits of 1 in the first binary data and the second binary data and the first total number of digits; or The first total number of bits is the total number of bits where the same bit is 0 in the first binary data and the second binary data, and the second total number of bits is the difference between the total number of bits of 0 in the first binary data and the second binary data and the first total number of bits.

3. The data processing method according to claim 1, characterized in that: The training data includes molecular structure data; The obtaining of n target data having the second feature comprises: The n target data with the second feature are obtained by searching for the largest common molecular substructure.

4. The data processing method according to claim 3, characterized in that: The step of obtaining n target data having the second characteristic by searching for the largest common molecular substructure comprises: Obtaining a candidate molecule dataset, wherein the application domain of the candidate molecule dataset is the same as the application domain of the training data; Determine p molecular graphs corresponding to at least part of the molecular data in the candidate molecular data set, where p is a positive integer greater than or equal to 0, and p is greater than or equal to n; Determining isomorphic parts of at least some of the p molecular graphs and the number of atoms in the isomorphic parts; Based on the similarity between the number of atoms of the candidate molecule in the candidate molecule data set and the number of atoms of the isomorphic part, n target data are determined from the candidate molecule data set.

5. A model training method, characterized in that: include: Obtaining a first target data set, wherein the first target data set includes target data obtained according to the method according to any one of claims 1 to 4, and the target data has no label information; The pre-trained model is trained using the first target data set to adjust model parameters of the pre-trained model to obtain an adjusted pre-trained model.

6. The model training method according to claim 5, characterized in that: Also includes: The adjusted pre-trained model is trained using training data of a specific application domain to adjust model parameters of the adjusted pre-trained model, wherein the training data of the specific application domain has label information, and the specific application domain is the same as or different from the application domain of the training data set.

7. The model training method according to claim 6, characterized in that: The step of training the adjusted pre-trained model using training data of a specific application domain comprises: Adding a linear classifier or a multilayer perceptron to the adjusted multiple pre-trained models to obtain one or more molecular prediction models or molecular evaluation models to be adjusted, wherein the adjusted multiple pre-trained models are trained according to the model training method according to claim 5 or 6; Using target data without label information to adjust model parameters of one or more molecular prediction models to be adjusted, to obtain one or more adjusted molecular prediction models; or, The target data with label information is used to adjust the model parameters of one or more molecular evaluation models to be adjusted, so as to obtain one or more adjusted molecular evaluation models.

8. The model training method according to claim 6, characterized in that: In the process of training the adjusted pre-trained model, at least one training task of context prediction and attribute masking is added.

9. A method for predicting molecular properties, characterized in that: include: Obtain molecular data; Based on the molecular data, a molecular prediction model obtained by using the adjusted pre-trained model provided according to any one of claims 5 to 8 is used to predict the molecular properties of the molecular data.

10. The molecular property prediction method according to claim 9, characterized in that: The molecular data has label information, and the method further includes: Based on the molecular data, a molecular evaluation model obtained by using the adjusted pre-training model is evaluated to obtain a molecular evaluation result of the molecular data.

11. A method for evaluating molecular properties, characterized in that: include: obtaining molecular data, wherein the molecular data has label information; Based on the molecular data, an evaluation is performed using a molecular evaluation model obtained by using the adjusted pre-training model provided according to any one of claims 5 to 8 to obtain a molecular evaluation result of the molecular data.

12. A design method, characterized in that: The method include: Obtaining molecular properties according to the method of claim 9, and / or obtaining molecular evaluation results according to the method of claim 11; Drug design or material design is performed based on the molecular properties and / or the molecular evaluation results.

13. A data processing device, characterized in that: include: A first feature acquisition module, used to obtain a first feature of m training data in a training data set, where m is an integer greater than or equal to 1; A target data acquisition module, used to obtain n target data with a second feature, so as to adjust the model parameters of the pre-trained model based on at least part of the n target data, where n is an integer greater than or equal to 1; wherein the first feature and the second feature meet the similarity requirement, the pre-trained model is trained based on the training data set, and the target data has no label information or has label information; The first feature includes a first fingerprint feature, and the second feature includes a second fingerprint feature; The similarity requirement includes: the similarity between the first fingerprint feature and the second fingerprint feature is greater than or equal to a preset similarity threshold; The first fingerprint feature includes first binary data, the second fingerprint feature includes second binary data, and each bit in the first binary data and the second binary data represents a substructure of a molecule; The obtaining of n target data having the second feature comprises: Obtaining a candidate data set, wherein the similarity between the second binary data of each candidate data in the candidate data set and the first binary data is greater than or equal to the preset similarity threshold, and the application domain of the candidate data set is the same as the application domain of the training data set; Obtain n candidate data from the candidate data set; or The first fingerprint feature includes first binary data, the second fingerprint feature includes second binary data, and each bit in the first binary data and the second binary data represents a substructure of a molecule; The obtaining of n target data having the second feature comprises: for each target data, Obtaining candidate data, wherein the application domain of the candidate data is the same as the application domain of the training data set; Obtain second binary data of a second fingerprint feature of the candidate data; If the similarity between the first binary data and the second binary data is greater than or equal to the preset similarity threshold, the candidate data is used as the target data.

14. A model training device, characterized in that: include: A data set acquisition module, configured to acquire a first target data set, wherein the first target data set includes target data acquired by the apparatus according to claim 13, and the target data has no label information; The model training module is used to train the pre-trained model using the first target data set to adjust the model parameters of the pre-trained model to obtain an adjusted pre-trained model.

15. A molecular property prediction device, characterized in that: include: A first molecular data acquisition module, used to obtain molecular data; A molecular property prediction module is used to predict the molecular properties of the molecular data based on the molecular data using a molecular prediction model obtained according to the adjusted pre-trained model provided according to claim 14.

16. A molecular property evaluation device, characterized in that: include: A second molecular data acquisition module, used to obtain molecular data, wherein the molecular data has label information; The molecular property evaluation module is used to evaluate the molecular data based on the molecular evaluation model obtained by the adjusted pre-training model provided according to claim 14 to obtain the molecular evaluation result of the molecular data.

17. A design device, characterized in that: include: An information acquisition module, used for obtaining the molecular property according to the molecular property prediction device according to claim 15, and / or obtaining the molecular evaluation result according to the molecular property evaluation device according to claim 16; A design module is used to perform drug design or material design based on the molecular properties and / or the molecular evaluation results.

18. An electronic device, characterized in that: include: processor; as well as A memory having executable codes stored thereon, which, when executed by the processor, causes the processor to execute the method according to any one of claims 1 to 12.

19. A computer-readable storage medium, characterized in that: An executable code is stored thereon, and when the executable code is executed by a processor of an electronic device, the processor is caused to execute the method according to any one of claims 1 to 12.

Images