Molecular data processing method and device, electronic equipment and storage medium

Abstract

The present disclosure provides a molecular data processing method and device, electronic equipment and storage medium, relates to the technical field of computers, in particular to the technical field of data processing, deep learning, artificial intelligence and the like. The specific implementation scheme is as follows: collecting from a source domain and a target domain to obtain a molecular data set; determining first original molecular features contained in the source domain and second original molecular features contained in the target domain from the molecular data set; determining model parameters based on the association relationship between the first original molecular features and the second original molecular features; and training based on the model parameters to generate a molecular property prediction model.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to the fields of data processing, deep learning, and artificial intelligence. Background Technology

[0002] Currently, few-shot learning is commonly used to process molecular data. However, this method neglects the characteristics of molecular property prediction, such as the different properties of the same molecule in different prediction tasks, and the relationships between molecules. Furthermore, in computer-aided molecular discovery, the distribution shift between the training and testing distributions can also affect the performance of molecular property prediction models. Summary of the Invention

[0003] This disclosure provides a method, apparatus, electronic device, and storage medium for processing molecular data.

[0004] According to one aspect of this disclosure, a method for processing molecular data is provided. The method may include: acquiring a molecular dataset from a source domain and a target domain; determining, from the molecular dataset, a first original molecular feature contained in the source domain and a second original molecular feature contained in the target domain, wherein the first original molecular feature is used to characterize the molecular properties of the molecular data in the source domain, and the second original molecular feature is used to characterize the molecular properties of the molecular data in the target domain; determining model parameters based on the correlation between the first and second original molecular features, wherein the correlation is updated as the feature values ​​of the first and / or second original molecular features change; and training based on the model parameters to generate a molecular property prediction model, wherein the molecular property prediction model is used to predict the molecular properties of the molecular data to be identified.

[0005] According to another aspect of this disclosure, a method for predicting molecular properties is also provided. This method may include: acquiring molecular data to be predicted; inputting the molecular data into a molecular property prediction model for prediction to obtain the molecular properties of the molecular data, wherein the molecular property prediction model is determined based on the aforementioned molecular data processing method.

[0006] According to one aspect of this disclosure, a molecular data processing apparatus is provided. The apparatus may include: a first acquisition unit for acquiring data from a source domain and a target domain to obtain a molecular dataset; a second acquisition unit for determining a first original molecular feature and a second original molecular feature from the molecular dataset, wherein the first original molecular feature characterizes the molecular properties of the molecular data in the source domain, and the second original molecular feature characterizes the molecular properties of the molecular data in the target domain; a determination unit for determining model parameters based on the correlation between the first original molecular feature and the second original molecular feature, wherein the correlation is updated as the feature values ​​of the first original molecular feature and / or the second original molecular feature change; and a generation unit for training based on the model parameters to generate a molecular property prediction model, wherein the molecular property prediction model is used to predict the molecular properties of the molecular data to be identified.

[0007] According to one aspect of this disclosure, a molecular property prediction apparatus is provided. The apparatus may include: a third acquisition unit for acquiring molecular data to be predicted; and a prediction unit for inputting the molecular data into a molecular property prediction model for prediction to obtain the molecular properties of the molecular data, wherein the molecular property prediction model is determined based on the aforementioned molecular data processing method.

[0008] According to another aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a molecular data processing method according to embodiments of this disclosure.

[0009] According to another aspect of this disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to perform a method for processing molecular data according to embodiments of this disclosure.

[0010] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements a method for processing molecular data according to embodiments of this disclosure.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0012] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0013] Figure 1This is a flowchart of a molecular data processing method according to an embodiment of the present disclosure;

[0014] Figure 2(a) is a flowchart of a method for predicting molecular properties according to an embodiment of the present disclosure;

[0015] Figure 2(b) is a schematic diagram of a computer product for performing a method for predicting molecular properties according to an embodiment of the present disclosure;

[0016] Figure 2(c) is a schematic diagram of a human-computer interaction interface for performing a method for predicting molecular properties according to an embodiment of the present disclosure;

[0017] Figure 3 This is a schematic diagram of predicting molecular tags in a related art according to an embodiment of the present disclosure;

[0018] Figure 4 This is a schematic diagram of a pre-trained model in a related art according to an embodiment of the present disclosure;

[0019] Figure 5 This is a schematic diagram illustrating how the relationship between molecules helps predict the properties of molecules according to an embodiment of the present disclosure.

[0020] Figure 6 This is a schematic diagram of a small-sample molecular property prediction scheme for domain migration according to an embodiment of the present disclosure;

[0021] Figure 7 This is a schematic diagram of a molecular data processing apparatus according to an embodiment of the present disclosure;

[0022] Figure 8 This is a schematic diagram of a molecular property prediction device according to an embodiment of the present disclosure;

[0023] Figure 9 This is a schematic diagram of an electronic device for a molecular data processing method according to an embodiment of the present disclosure. Detailed Implementation

[0024] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0025] The following describes a method for processing molecular data according to an embodiment of this disclosure.

[0026] Figure 1This is a flowchart of a molecular data processing method according to an embodiment of the present disclosure, such as... Figure 1 As shown, the method may include the following steps:

[0027] Step S102: Collect data from the source domain and the target domain to obtain a molecular dataset.

[0028] In the technical solution provided by step S102 of this disclosure, the source domain can be a training sample dataset, the target domain can be a test sample dataset, and the molecular dataset can include molecular data randomly sampled from the auxiliary basic dataset of the source domain and molecular data randomly sampled from the molecular property prediction task set of the target domain.

[0029] In this embodiment, a joint sampling method can be used to simultaneously sample randomly from the source domain and the target domain to obtain a molecular dataset. Compared with the method in related technologies that only samples from the target domain to obtain a molecular dataset, this solves the problem of distribution shift between the training distribution and the test distribution, thereby achieving the technical effect of improving the accuracy of the model in predicting molecular properties.

[0030] Optionally, the aforementioned molecular dataset can also be a set of meta-training task pairs. The molecular data randomly sampled from the auxiliary base dataset of the source domain can be used as the source domain task, and the molecular data randomly sampled from the molecular property prediction task set of the target domain can be used as the target domain task. The source domain task and the target domain task can form a meta-training task pair, and sampling can be repeated multiple times to obtain the molecular dataset.

[0031] For example, the auxiliary underlying dataset of the source domain is From random sampling As the source domain task, the target domain molecular property prediction task set is Z. l It can be from Z l random sampling As the target domain task, the meta-training task pair can be: Molecular datasets can be:

[0032] It should be noted that the above-described method for collecting molecular datasets from the source and target domains is only a preferred embodiment of this disclosure. The method for collecting molecular datasets from the source and target domains is not specifically limited here. Any method and process for collecting molecular datasets from the source and target domains is within the protection scope of this disclosure, and will not be listed here.

[0033] Step S104: Determine the first original molecular features contained in the source domain and the second original molecular features contained in the target domain from the molecular dataset.

[0034] In the technical solution provided in step S104 of this disclosure, the molecular dataset in the source domain may include a source domain support dataset and a source domain query dataset, and the molecular dataset in the target domain may include a target domain support dataset and a target domain query dataset. The source domain support dataset and the target domain support dataset can be used to train a molecular property prediction model, and the source domain query dataset and the target domain query dataset can be used to test the molecular property prediction model. The first original molecular feature can be a property-aware molecular representation of randomly sampled molecular data in the source domain, and the second original molecular feature can be a property-aware molecular representation of randomly sampled molecular data in the target domain. The molecular representation can be a measure of a molecule's properties in a certain aspect, which can be either the physicochemical properties of the molecule or a numerical index describing molecular structure information derived from the molecular structure through various algorithms. The property-aware molecular representation can be used to characterize the molecular properties of the molecule. That is, the first original molecular feature can be used to characterize the molecular properties of the molecular data in the source domain, and the second original molecular feature can be used to characterize the molecular properties of the molecular data in the target domain.

[0035] In this embodiment, a graph-based molecular encoder can be used to determine the first original molecular features contained in the source domain and the second original molecular features contained in the target domain from the molecular dataset. The molecular encoder can be a pre-trained graph neural network (pre-GNN), which is only an example and is not specifically limited.

[0036] For example, we can first target the source domain tasks mentioned above. support set and query set The molecule x in s,i Extraction length d g Molecular characterization g s,i , where d g This represents a hyperparameter, and a suitable value can be found through parameter tuning; then, based on the extracted g... s,i Calculate the class representation of active and inactive. The formula for calculating can be shown below:

[0037]

[0038] In the above formula, c can be 0 or 1, y s,i It can be molecule x s,i Property labels, It can be molecule x s,i And its active binary set.

[0039] Next, the class representatives obtained above can be used as... As semantic information, it is further encoded using a scaled dot product self-attention mechanism to obtain the encoded b. s,i b s,i The formula for calculating can be shown below:

[0040]

[0041] In the above formula, C s,i T It is C s,i The transpose of , where,

[0042] Finally, the first original molecular feature can be obtained through the following formula, that is, the property-sensing molecular characterization p of randomly sampled molecular data from the source domain. s,i It can be as follows:

[0043]

[0044] in, It can represent being W p Parameterized multilayer perceptron, W p The number of layers in the MLP depends on the number of layers, which can be determined experimentally; that is, how many layers does the MLP have, and W... p This includes the weight matrices of several fully connected layers. It should be noted that the method described above for calculating the first original molecular feature can also be used to calculate the second original molecular feature; however, this will not be elaborated upon here.

[0045] It should be noted that the above-mentioned calculation method for the first and second original molecular features is only a preferred embodiment. The calculation method for the first and second original molecular features is not specifically limited here. Any calculation method and process for the first and second original molecular features is within the protection scope of the embodiments of this disclosure, and will not be described in detail here.

[0046] Step S106: Determine the model parameters based on the correlation between the first and second original molecular features.

[0047] In the technical solution provided by step S106 of this disclosure, the association relationship may include a dependency graph and a joint relationship graph between the first original molecular feature and the second original molecular feature, which can be updated as the feature values ​​of the first original molecular feature and / or the feature values ​​of the second original molecular feature change. The model parameters may be meta-training parameters used for meta-training of molecular data. The model parameters may include meta-training parameters learned from specific meta-task pairs, wherein the meta-training parameters can be used to learn the characteristics of the task pairs. The model parameters may also include meta-training parameters learned from a large set of small sample task pairs, wherein the meta-training parameters can be used to learn the commonalities of the task pairs.

[0048] In this embodiment, the meta-training can be divided into inner loop training and outer loop training. The inner loop training loss function and the outer loop training loss function can be determined according to the correlation between the first original molecular feature and the second original molecular feature. Then, the model parameters are obtained by combining the inner loop training loss function and the outer loop training loss function with all meta-training task pairs in the task pair set.

[0049] Optionally, the inner loop training loss function can be calculated using the loss function on the source domain support dataset and the loss function on the target domain support dataset, and the outer loop training loss function can be calculated using the loss function on the source domain query dataset and the loss function on the target domain query dataset. Here, no specific limitations are imposed on the calculation methods of the inner loop training loss function and the outer loop training loss function.

[0050] Step S108: Train the model based on the model parameters to generate a molecular property prediction model.

[0051] In the technical solution provided by step S108 of this disclosure, a molecular property prediction model for predicting the molecular properties of molecular data to be identified can be generated by training the model parameters calculated in step S106. This achieves the technical effect of quickly predicting molecular properties in small sample molecular property prediction tasks involving domain transfer. The molecular property prediction model can be a small sample molecular property prediction model for domain transfer (Transferable Property-Aware Relation Networks, abbreviated as T-PAR).

[0052] Through steps S102 to S108, molecular datasets are collected from the source and target domains. First, the first original molecular features in the source domain and the second original molecular features in the target domain are determined from the molecular datasets. Model parameters are determined based on the correlation between the first and second original molecular features. Finally, a molecular property prediction model is generated by training the model based on these parameters. In other words, the first original molecular features of the source domain and the second original molecular features of the target domain are first determined from the molecular datasets collected from the source and target domains. Then, model parameters are determined based on the correlation between the first and second original molecular features. Finally, the molecular property prediction model is obtained by training the model parameters. This avoids distribution shifts between the training and testing distributions, thereby improving the model performance of the molecular property prediction model. This solves the technical problem of low accuracy in molecular property prediction and achieves the technical effect of improving the accuracy of molecular property prediction.

[0053] The method described in this embodiment will now be described in further detail.

[0054] As an optional implementation, step S106, determining the model parameters based on the correlation between the first original molecular feature and the second original molecular feature, includes: updating the first original molecular feature to the first target molecular feature based on the correlation, and updating the second original molecular feature to the second target molecular feature; and obtaining the model parameters based on the first target molecular feature and the second target molecular feature.

[0055] In this embodiment, the property-aware representation in the correlation relationship between the first original molecular feature and the second original molecular feature can be iteratively updated. The first original molecular feature can be updated to the first target molecular feature, and the second original molecular feature can be updated to the second target molecular feature. Then, model parameters are obtained based on the first target molecular feature and the second target molecular feature to improve the model performance of the property prediction model constructed based on the model parameters. This achieves the technical effect of improving the accuracy of the model in predicting molecular properties. The first target molecular feature can be the molecular representation after iteratively updating molecular data randomly sampled from the source domain, and the second target molecular feature can be the molecular representation after iteratively updating molecular data randomly sampled from the target domain.

[0056] Optionally, before updating the first original molecular feature to the first target molecular feature and the second original molecular feature to the second target molecular feature, the molecular representations in the first original molecular feature and the second original molecular feature can be initialized respectively.

[0057] For example, the set of all molecules included in the association between the first and second original molecular features is identified as follows: V s,m V is the set of all molecules contained in the first primitive molecular characteristic. t,m V is the set of all molecules contained in the second primitive molecular characteristic. s,m The p-th molecule x s,p The initial molecular characterization is as follows x ω,p The molecular characterization during the τth iteration is as follows V t,m The p-th molecule x t,p The initial molecular characterization is as follows x ω,p The molecular characterization during the τth iteration is as follows

[0058] It should be noted that the above-described method for updating the first original molecular feature to the first target molecular feature and the second original molecular feature to the second target molecular feature is merely an example. The method for updating the first original molecular feature to the first target molecular feature and the second original molecular feature to the second target molecular feature is not specifically limited here. Any method and process for updating the first original molecular feature to the first target molecular feature and the second original molecular feature to the second target molecular feature is within the protection scope of the embodiments of this disclosure, and will not be described in detail here.

[0059] As an optional implementation, the method further includes: constructing a relationship based on the source domain supporting dataset, the source domain query dataset, the target domain supporting dataset, and the target domain query dataset.

[0060] In this embodiment, the molecular dataset in the source domain may include a source domain support dataset and a source domain query dataset, and the molecular dataset in the target domain may include a target domain support dataset and a target domain query dataset. The source domain support dataset and the target domain support dataset can be used to train a molecular property prediction model, and the source domain query dataset and the target domain query dataset can be used to test the molecular property prediction model.

[0061] In this embodiment, the association between the first original molecular feature and the second original molecular feature can be constructed based on the source domain support dataset and source domain query dataset included in the molecular dataset in the source domain, and the target domain support dataset and target domain query dataset included in the molecular dataset in the target domain. A joint relationship graph corresponding to the association can also be constructed. Then, the model parameters are determined based on the joint relationship graph to improve the model performance of the property prediction model constructed based on the model parameters, thereby achieving the technical effect of improving the accuracy of the model in predicting molecular properties.

[0062] It should be noted that the above method for constructing the correlation between the first original molecular feature and the second original molecular feature is only a preferred embodiment. The method for constructing the correlation between the first original molecular feature and the second original molecular feature is not specifically limited here. Any method and process for constructing the correlation between the first original molecular feature and the second original molecular feature is within the protection scope of the embodiments of this disclosure, and will not be described in detail here.

[0063] As an optional implementation method, a relationship is constructed based on the source domain supporting dataset, the source domain query dataset, the target domain supporting dataset, and the target domain query dataset, including: constructing a relationship based on the source domain supporting dataset, a portion of the source domain query dataset in the source domain query dataset, the target domain supporting dataset, and a portion of the target domain query dataset in the target domain query dataset.

[0064] In this embodiment, a portion of the source domain query dataset can be a molecule randomly drawn from the source domain query dataset, for example, from the source domain query dataset. The molecule (x) randomly selected from the sample s,j1 ,y s,j1 The target domain query dataset can be a subset of the target domain query dataset that is randomly drawn from the target domain query dataset. For example, from the target domain query dataset... The molecule (x) randomly selected from the sample t,j2 ,y t,j2 ).

[0065] In this embodiment, a correlation between a first original molecular feature and a second original molecular feature can be constructed based on all molecules in the source domain support dataset, a portion of molecules in the source domain query dataset, all molecules in the target domain support dataset, and a portion of molecules in the target domain query dataset. A property prediction model can then be constructed based on this correlation to improve the accuracy of the model in predicting molecular properties.

[0066] It should be noted that the above method for constructing a relationship based on the source domain support dataset, a portion of the source domain query dataset, the target domain support dataset, and a portion of the target domain query dataset is only a preferred embodiment. No specific limitations are imposed on the method for constructing a relationship based on the source domain support dataset, a portion of the source domain query dataset, the target domain support dataset, and a portion of the target domain query dataset. Any method and process for constructing a relationship based on the source domain support dataset, a portion of the source domain query dataset, the target domain support dataset, and a portion of the target domain query dataset is within the protection scope of this disclosure, and will not be elaborated upon here.

[0067] As an optional implementation, a relationship is constructed based on the source domain support dataset, a portion of the source domain query dataset, the target domain support dataset, and a portion of the target domain query dataset. This includes: obtaining the union of the source domain support dataset, the portion of the source domain query dataset, the target domain support dataset, and the portion of the target domain query dataset; and obtaining and concatenating the original similarity between molecular properties of molecular data in the same domain and / or the original similarity between molecular properties of molecular data in different domains.

[0068] In this embodiment, the aforementioned association may further include the similarity between molecular properties of molecular data in the same domain and / or the similarity between molecular properties of molecular data in different domains, wherein the same domain may include the source domain or the target domain, and different domains may include the source domain and the target domain.

[0069] In this embodiment, the union of the source domain support dataset, a portion of the source domain query dataset in the source domain query dataset, the target domain support dataset, and a portion of the target domain query dataset in the target domain query dataset can be obtained first. Then, the original similarity between molecular properties of molecular data in the same domain and / or the original similarity between molecular properties of molecular data in different domains in the union can be calculated. Finally, the association relationship can be constructed based on the calculation results.

[0070] For example, obtaining the union of the source domain support dataset, a portion of the source domain query dataset from the source domain query dataset, the target domain support dataset, and a portion of the target domain query dataset from the target domain query dataset can be identified as the set of all molecules contained in the m-th union graph. in, V s,m This is the union of the source domain support dataset and a portion of the source domain query dataset. To support the dataset for the source domain, (x s,j1,y s,j1 V represents a portion of the source domain query dataset. t,m To provide the union between the target domain support dataset and a portion of the target domain query dataset, To support the dataset for the target domain, (x t,j2 ,y t,j2 ) represents a portion of the target domain query dataset.

[0071] Optionally, the original similarity between molecular properties of molecular data in the same domain and the original similarity between molecular properties of molecular data in different domains can also be represented by an adjacency matrix, for example:

[0072]

[0073] in, It can be used to represent dense matrices, that is, matrices where every position has a value. and It can be used to represent the original similarity between molecular properties of molecular data in the same domain, that is, This represents the original similarity between molecular properties in the source domain, where ss represents the intermolecular similarity between molecules in the source domain. This represents the original similarity between molecular properties of molecular data in the target domain, where tt represents the intermolecular similarity within the target domain. and It can be used to represent the original similarity between molecular properties of molecular data in different domains, that is, the original similarity between molecular properties of molecular data in the source domain and molecular properties of molecular data in the target domain. st and ts represent the molecular similarity between the source domain and the target domain.

[0074] Optionally, the above and It can also be used to control the propagation of molecular characterization within the same domain, as mentioned above. and It can also be used to control the propagation of molecular characterization in different domains, without specific limitations here.

[0075] For example, the above The similarity between the p-th molecule and the q-th molecule can be corresponding to The middle position is (p, q) if It refers to the original similarity between molecular properties in molecular data within the same domain, that is, exist and In this case, it can be calculated using the following formula:

[0076]

[0077] In the above formula, W a It can be used to represent parameters of a multilayer perceptron.

[0078] Optionally, if the above It refers to the original similarity between molecular properties in molecular data from different domains, that is, and In the middle, it is possible to change the parameter W a Change the parameter to W b The calculations yielded the results.

[0079] It should be noted that the methods for representing and calculating the original similarity between molecular properties of molecular data in the same domain and between molecular properties of molecular data in different domains are merely examples. No specific limitations are imposed on the methods for representing and calculating the original similarity between molecular properties of molecular data in the same domain and between molecular properties of molecular data in different domains. Any method, calculation method, or process used to represent the original similarity between molecular properties of molecular data in the same domain and between molecular properties of molecular data in different domains is within the protection scope of the embodiments of this disclosure, and will not be described in detail here.

[0080] As an optional implementation, the method further includes: sparsifying the original similarity between molecular properties of molecular data in the same domain of the union to obtain a target similarity between molecular data in the same domain of the union, wherein the target similarity is greater than a first similarity threshold; and / or sparsifying the original similarity between molecular properties of molecular data in different domains of the union to obtain a target similarity between molecular data in different domains of the union, wherein the target similarity is greater than a second similarity threshold.

[0081] In this embodiment, the original similarity between molecular properties in molecular data within the same domain and / or between molecular properties in molecular data across different domains obtained through the above steps can be sparsified to obtain and concentrate the target similarity between molecular data within the same domain and / or concentrate the target similarity between molecular data across different domains. This achieves the goal of retaining only the edges between similar molecules, ensuring that the update of each molecular representation is only affected by molecules that are very similar to it. This achieves the technical effect of removing the influence of noise, accelerating information dissemination, and reducing the amount of computation. Specifically, the target similarity between molecular properties in molecular data within the same domain is greater than a first similarity threshold, and the target similarity between molecular properties in molecular data across different domains is greater than a second similarity threshold.

[0082] For example, regarding the above and When performing sparsification, you can use As a threshold, where q can represent the q-th column vector in the extraction matrix, if the value is greater than the threshold, the original value is retained; if it is less than the threshold, it is set to 0. And for the above... When performing sparsity reduction, you can use... As a threshold, in the and After performing sparsification, the above target similarity can be obtained:

[0083]

[0084] in, To Similarity after sparsification To Similarity after sparsification To Similarity after sparsification To Similarity after sparsification.

[0085] It should be noted that the methods described above for calculating the target similarity between molecular properties of molecular data in the same domain and between molecular properties of molecular data in different domains are merely examples. No specific limitations are imposed on the methods for calculating the target similarity between molecular properties of molecular data in the same domain and between molecular properties of molecular data in different domains. Any method and process used to calculate the target similarity between molecular properties of molecular data in the same domain and between molecular properties of molecular data in different domains are within the protection scope of the embodiments of this disclosure, and will not be described in detail here.

[0086] As an optional implementation, step S106, based on the correlation between the first original molecular features and the second original molecular features, determines the model parameters, including: determining the loss function on the source domain support dataset, the loss function on the source domain query dataset, the loss function on the target domain support dataset, and the loss function on the target domain query dataset based on the correlation; and determining the model parameters based on the loss functions on the source domain support dataset, the source domain query dataset, the target domain support dataset, and the target domain query dataset.

[0087] In this embodiment, the loss function on the source domain support dataset, the loss function on the source domain query dataset, the loss function on the target domain support dataset, and the loss function on the target domain query dataset can be determined by querying the correlation between the first original molecular feature and the second original molecular feature. The model parameters are then determined based on the obtained multiple loss functions, and the model parameters are trained to generate a molecular property prediction model. This achieves the technical effect of improving the accuracy of the molecular property prediction model in predicting molecular properties.

[0088] For example, the loss function on the source domain supporting dataset. It can be calculated using the following formula:

[0089]

[0090] In the above formula, if V s,m If the true labels of the p-th molecule and the q-th molecule are the same (e.g., both are 0 or both are 1), then G s,m The truth value of a relation graph can be defined as [G] s,m ] pq =1; if V s,m If the true labels of the p-th molecule and the q-th molecule are not the same (e.g., the p-th molecule is 0 and the q-th molecule is 1), then G... s,m The truth value of a relational graph can be defined as [G] s,m ] pq =0.

[0091] Loss function on source domain query dataset It can be calculated using the following formula:

[0092]

[0093] It should be noted that the loss function used above for calculating the source domain support dataset is... loss function on the source domain query dataset The calculation formula can also be used to calculate the loss function on the target domain support dataset. and the objective function on the source domain query dataset I will not go into detail here.

[0094] It should be noted that the above methods for calculating the loss function on the source domain supporting dataset, the source domain query dataset, the target domain supporting dataset, and the target domain query dataset are merely preferred embodiments. No specific limitations are imposed on the methods for calculating these functions. Any methods and processes used to calculate the loss function on the source domain supporting dataset, the source domain query dataset, the target domain supporting dataset, and the target domain query dataset are within the protection scope of this disclosure. Examples will not be provided here.

[0095] As an optional implementation, model parameters are determined based on the loss function on the source domain supporting dataset, the loss function on the source domain query dataset, the loss function on the target domain supporting dataset, and the loss function on the target domain query dataset. This includes: determining first model parameters based on the loss function on the source domain supporting dataset and the loss function on the target domain supporting dataset; and determining second model parameters based on the loss function on the source domain query dataset, the loss function on the target domain query dataset, and the first model parameters.

[0096] In this embodiment, the model parameters can be meta-training parameters, which may include first model parameters and second model parameters. The first model parameters can be meta-training parameters learned from specific meta-task pairs and can be used to learn the characteristics of task pairs. The second model parameters can be meta-training parameters learned from a large set of small sample task pairs and can be used to learn the commonalities of multiple task pairs. The task pairs can be meta-training task pairs, including molecular data in the source domain and molecular data in the target domain.

[0097] In this embodiment, meta-training can be divided into inner loop training and outer loop training. First, the inner loop training loss function is determined using the loss functions on the source domain support dataset and the target domain support dataset. Then, the first model parameters are obtained by combining the inner loop training loss function with all meta-training task pairs in the task pair set. Similarly, the outer loop training loss function is determined by first using the loss functions on the source domain query dataset and the target domain query dataset. Then, the second model parameters are obtained by combining the outer loop training loss function with all meta-training task pairs in the task pair set. Finally, the first and second model parameters are trained to generate a molecular property prediction model, thereby improving the accuracy of the molecular property prediction model in predicting molecular properties.

[0098] Alternatively, the above inner loop training loss function can be calculated using the following formula, for example:

[0099]

[0100] In the above formula, if V s,m If the true labels of the p-th molecule and the q-th molecule are the same (e.g., both are 0 or both are 1), then... The truth value of a relation graph can be defined as If V s,m If the true labels of the p-th molecule and the q-th molecule are not the same (e.g., the p-th molecule is 0 and the q-th molecule is 1), then... The truth value of a relation graph can be defined as It should be noted that, It can also be used with The same method is used for definition, which will not be elaborated here.

[0101] Optionally, based on the above formula, it can be... Updated to target meta-training task pairs The task And update the general information in the inner loop.

[0102] Alternatively, the outer loop training loss function described above can be calculated using the following formula, for example:

[0103]

[0104] Based on the above formula, and by combining all task pairs in the meta-training task pair set, the model parameters in this embodiment can be obtained.

[0105] It should be noted that the above method for calculating model parameters is only a preferred embodiment. The method for calculating model parameters is not specifically limited here. Any method and process used to calculate model parameters is within the protection scope of the embodiments of this disclosure, and will not be described in detail here.

[0106] As an optional implementation, step S102, collecting data from the source domain and the target domain to obtain a molecular dataset, includes: randomly sampling from the source domain to obtain molecular data in the source domain, and randomly sampling from the target domain to obtain molecular data in the target domain.

[0107] In this embodiment, a joint sampling method can be used to randomly sample from the source domain to obtain molecular data in the source domain, and randomly sample from the target domain to obtain molecular data in the target domain. That is, random sampling is performed simultaneously from both the source and target domains to obtain a molecular dataset. Compared with the method in related technologies that only samples from the target domain to obtain a molecular dataset, this solves the problem of distribution offset between the training and testing distributions, and achieves the technical effect of improving the accuracy of the model in predicting molecular properties.

[0108] It should be noted that the above-described method for collecting molecular datasets from the source and target domains is only a preferred embodiment of this disclosure. The method for collecting molecular datasets from the source and target domains is not specifically limited here. Any method and process for collecting molecular datasets from the source and target domains is within the protection scope of this disclosure, and will not be listed here.

[0109] The method for predicting molecular properties according to embodiments of this disclosure will be described below.

[0110] Figure 2(a) is a flowchart of a method for predicting molecular properties according to an embodiment of the present disclosure. As shown in Figure 2(a), the method may include the following steps:

[0111] Step S202: Obtain the molecular data to be predicted.

[0112] In the technical solution provided by step S202 of this disclosure, the molecular data to be predicted can be molecular data to be identified, such as new drugs discovered in actual drug discovery applications. It should be noted that this is merely an example and does not specifically limit the type of molecular data to be predicted. Any type of molecular data to be predicted is within the protection scope of the embodiments of this disclosure, and will not be listed here.

[0113] Step S204: Input the molecular data into the molecular property prediction model for prediction to obtain the molecular properties of the molecular data.

[0114] In the technical solution provided by step S204 of this disclosure, the molecular property prediction model first collects molecular datasets from the source domain and the target domain, then determines the model parameters based on the correlation between the first original molecular features contained in the source domain and the second original molecular features contained in the target domain determined from the molecular dataset, and finally trains the model parameters to obtain the model. This molecular property prediction model can be used to predict the molecular properties of the molecular data to be identified.

[0115] In this embodiment, the molecular properties of the molecular data can be obtained by inputting the molecular data to be predicted into the molecular property prediction model, thereby improving the accuracy of the model in predicting molecular properties.

[0116] Figure 2(b) is a schematic diagram of a computer product for performing a method for predicting molecular properties according to an embodiment of the present disclosure. As shown in Figure 2(b), the above steps S202 to S204 can be performed by a computing device, the molecular data is input into the computing device, the computing device predicts the molecular properties of the molecular data based on the molecular property prediction model, and outputs the molecular properties of the molecular data, so as to achieve the technical effect of improving the accuracy of the model in predicting molecular properties.

[0117] Figure 2(c) is a schematic diagram of a human-computer interaction interface for performing a method for predicting molecular properties according to an embodiment of the present disclosure. As shown in Figure 2(c), the molecular data 202 to be predicted can be input by triggering the control 201 on the interface, and the molecular properties 203 of the molecular data to be predicted can be obtained.

[0118] It should be noted that no specific limitations are made here on the computer products and human-computer interaction interfaces used to perform methods for predicting molecular properties. Any computer products and human-computer interaction interfaces used to perform methods for predicting molecular properties are within the scope of protection of this disclosure, and will not be listed here.

[0119] Through the above steps S202 to S204, the molecular data to be predicted is obtained, and the molecular data is input into the molecular property prediction model for prediction to obtain the molecular properties of the molecular data. This solves the technical problem of low accuracy of the model in predicting molecular properties and achieves the technical effect of improving the accuracy of the model in predicting molecular properties.

[0120] The molecular data processing method of this disclosure will be further described below with reference to preferred embodiments.

[0121] Molecular property prediction is one of the most critical tasks in the computer-aided drug discovery process, playing a vital role in many downstream applications, such as drug screening and drug design.

[0122] Currently, molecular property prediction is mainly accomplished by treating molecules as graph data and combining it with deep learning. However, in the process of new drug development, the accuracy of molecular property prediction is low due to the limited number of labeled samples (e.g., molecules with known pharmacological properties).

[0123] In related technologies, molecular property prediction methods mostly follow a fixed framework: the molecule is regarded as a graph, the atoms are the nodes in the graph, the chemical bonds are the edges, the molecule is fed into a graph neural network (GNN) to obtain the molecular representation (graph representation), and then the molecular property prediction results are obtained through a few-sample learning method.

[0124] Figure 3 This is a schematic diagram of predicting molecular tags in a related art according to an embodiment of the present disclosure, such as... Figure 3 As shown, g, g', f, f', and k are the embedding models. The compounds in the dataset include lithium ions, ethanol, caffeine, toluenesulfonic acid, and dopamine. Lithium ions and toluenesulfonic acid are labeled as active, while ethanol, caffeine, and dopamine are labeled as inactive. The novel compound to be predicted is styrene oxide. The Iterative Refinement Long Short-Term Memory (IterRefLSTM) neural network can predict the label of styrene oxide by multiplying the similarity between the molecular representations of the compounds in the dataset and the novel compound with the labels of the compounds in the dataset and then summing the results.

[0125] Figure 4 This is a schematic diagram of a pre-trained model in a related art according to an embodiment of the present disclosure, such as... Figure 4 As shown, the Meta-Molecular Graph Neural Network (Meta-MGNN) model based on meta-learning combines the classic algorithm in few-shot learning (Model-Agnostic Meta-Learning, MAML) with pre-GNN, and uses the molecular representation pre-training model provided in pre-GNN for training.

[0126] Figure 5 This is a schematic diagram illustrating how the relationship between molecules helps predict the properties of molecules according to an embodiment of the present disclosure. Figure 5 As shown, Property-Aware Relation Networks (PARs) consider the different performances of the same molecule in different property prediction tasks, and that the relationships between molecules change with different properties. Two molecules with common properties may also differ in another property. Therefore, the relationships between molecules can also be used to assist in predicting the properties of molecules.

[0127] However, in the embodiments of this disclosure, a small-sample molecular property prediction method T-PAR for domain transfer is proposed. This method takes into account the distribution shift that may occur during training and testing in actual molecular discovery applications, and solves the problem of transferable small-sample molecular property prediction through joint sampling and joint relation graph learning strategies, so as to improve the technical effect of improving the accuracy of molecular property prediction.

[0128] The following is a further description of the small-sample molecular property prediction method for domain migration in the embodiments of this disclosure.

[0129] Figure 6 This is a schematic diagram of a small-sample molecular property prediction scheme for domain migration according to an embodiment of the present disclosure, as shown below. Figure 6 As shown, through joint sampling, meta-training task pairs can be obtained simultaneously from both the source and target domains, and T is sampled from the auxiliary base dataset D from the source domain. s As a source domain task, the task set Z predicts molecular properties. l random sampling As the target domain task, the two can form a meta-training task pair. Repeat the above process until the preset maximum number of samples is reached. The number of tasks to obtain a task pair set.

[0130] In this embodiment of the disclosure, graph-based molecular encoders (e.g., GIN and Pre-GNN, assuming parameters W) can also be used. g ),for support set and query set The molecule x s,i Extraction length d g The characterization of g s,i , where d g It is a hyperparameter that can be used to find a suitable value through parameter tuning to support the set. It can contain K labeled samples from each class as training data. Query set It can contain M samples for testing. It should be noted that K is generally small, therefore this is a few-shot learning task.

[0131] Optionally, meta-training can be divided into an outer loop and an inner loop. In the outer loop, the meta-training parameters are learned from a large set of task pairs, which is equivalent to capturing the commonalities of these task pairs; in the inner loop, the focus is on a specific task. support set It serves the same purpose as the training set in regular training, used to fine-tune the meta-training parameters to the current task. query set It serves as a test set in ordinary training, used to test the predictive performance of the model parameters after fine-tuning.

[0132] In this embodiment of the disclosure, for The active and inactive class representatives can be calculated as follows: in, y is a binary set of molecules and their activities. s,i It is molecule x s,i The property label, c can be 0 or 1.

[0133] The above class representatives are used as semantic information and further encoded using a scaled dot product self-attention mechanism: in, C s,i T It is C s,i The transpose of .

[0134] Optionally, the class representative can be an average characterization of molecules that are currently active and inactive.

[0135] like Figure 6 As shown in the embodiments of this disclosure, property-aware molecular characterization can also be calculated: in, Indicated by W p Parameterized multilayer perceptron, W p The parameters depend on the number of layers in the MLP and can be determined experimentally; that is, the number of layers in the MLP determines the parameters W. p The parameters can include the weight matrices of several fully connected layers.

[0136] It should be noted that for the task set Z predicting from molecular properties l random sampling Alternatively, the above calculation method can be used to obtain the property-sensing molecular characterization p. t,i .

[0137] In this embodiment of the disclosure, a joint relationship graph also needs to be constructed. The m-th joint relationship graph may include: source domain support set. All molecules, source domain query sets A randomly selected molecule (x) s,j1 ,y s,j1 ), target domain support set All molecules in, and source domain query set A randomly selected molecule (x) t,j2 ,y t,j2 )

[0138] like Figure 6As shown, the set of all molecules included in the m-th joint relationship graph is labeled as follows: in, It should be noted that for each All node representations will be iteratively updated T times on the joint relationship graph.

[0139] Optionally, V s,m The p-th molecule x s,p The representation Initialize to the corresponding property-aware representation p s,p ,use x represents ω,p Molecular characterization during the τth iteration.

[0140] In this embodiment of the disclosure, the adjacency matrix can also be estimated using the current molecular characterization:

[0141]

[0142] in, and Controlling the propagation of molecular characterization within the same task, and Controlling the propagation of molecular characterization across different tasks, ss is targeted at V s,m Intermolecular in ... t,m Intermolecular in ... s,m With V t,m Intermolecular distance.

[0143] Optionally, The similarity correspondence between the p-th molecule and the q-th molecule The (p, q) position in It can be done and Calculate, if exist and In this case, it can be calculated using the following formula:

[0144]

[0145] Among them, W a These are the parameters of the multilayer perceptron.

[0146] Optionally, if exist and In the middle, it is possible to change the parameter W a Change the parameter to W b The calculations yielded the results.

[0147] It should be noted that the results obtained through the above method It is a dense matrix, meaning that each position has a value.

[0148] In this disclosed embodiment, it is also necessary to further... The purpose of sparsification is to retain only the edges between similar molecules, so that the update of each molecule's representation is only affected by molecules that are very similar to it, thereby removing the influence of noise, accelerating the spread of information, and reducing the amount of computation.

[0149] because The domain contains both molecular relationships within the same domain and relationships between molecules from different domains; therefore, different thresholds can be used to differentiate them. and Processing, for example, the above and When performing sparsification, you can use As a threshold, where q can represent the q-th column vector in the extraction matrix, if the value is greater than the threshold, the original value is retained; if it is less than the threshold, it is set to 0. And for the above... When performing sparsity reduction, you can use... As a threshold, after sparsification, we can obtain

[0150] In After sparsification, it is possible to Update the node representation on the corresponding sparsed joint graph, and V s,m The characterization of all molecules is uniformly placed into the characterization matrix. Where the p-th column corresponds to Similarly, V can also be used. t,m The characterization of all molecules is uniformly placed into the characterization matrix.

[0151] Alternatively, it can also be The node representations are uniformly placed into the representation matrix. It was updated to Then after T iterations, we get and And this is used as the final molecular characterization, in which, and To ultimately optimize the relationship graph, W r These are parameters that can be learned.

[0152] Next, for each task, the molecule x s,p and x t,p Perform property prediction: z s，m，p =softmax(Wc ·h s，m，p ) and z t，m，p =softmax(W c ·h t，m，p ), where W c For parameters.

[0153] The T-PAR model can be expressed as: in, Including parameters used to learn the commonalities of the task, This includes parameters used to learn the characteristics of the task.

[0154] In this embodiment of the disclosure, a loss function on the support set can also be designed by querying the dependency graph. This loss function can be:

[0155]

[0156] If V s,m If the true labels of the p-th molecule and the q-th molecule are the same (e.g., the true labels of the p-th molecule and the q-th molecule are both 0 or both 1), then G s,m The truth value of a relational graph can be defined as [G] s,m ] pq =1, conversely, G s,m The truth value of a relational graph can be defined as [G] s,m ] pq =0. It should be noted that the loss function for the target domain can also be calculated using the above method. I will not go into details here.

[0157] Meanwhile, in this embodiment, both the source domain and the target domain need to be considered to calculate the inner loop training loss, and the loss function can be:

[0158]

[0159] If V s,m If the true labels of the p-th molecule and the q-th molecule are the same (e.g., the true labels of the p-th molecule and the q-th molecule are both 0 or both 1), then G s,m The truth value of a relation graph can be defined as Conversely, G s,m The truth value of a relation graph can be defined as

[0160] In this embodiment of the disclosure, a loss function on the query set can also be designed through a joint relationship graph. This loss function can be:

[0161]

[0162] It should be noted that the loss function for the target domain can also be calculated using the above method. I will not go into details here.

[0163] In this embodiment of the disclosure, the outer loop training loss can also be calculated using the following formula, for example: For sets Meta-training parameters can also be obtained for all task pairs. and

[0164] It should be noted that each task in the above inner loop is a small sample molecular property prediction task. The above embodiments of this disclosure are to learn how to quickly solve a new small sample molecular property prediction task from a large number of small sample molecular property prediction tasks.

[0165] In the embodiments of this disclosure, the physicochemical properties of molecules can be predicted quickly through joint sampling and joint relation graph learning strategies, thereby assisting in finding compounds that meet the expected properties from a large number of candidate compounds. This not only speeds up molecular screening and molecular design, but can also be widely applied in industrial steps such as virtual screening and lead compound optimization, thereby solving the technical problem of low accuracy in molecular property prediction and achieving the technical effect of improving the accuracy of molecular property prediction.

[0166] This disclosure also provides an embodiment for performing Figure 1 The molecular data processing method apparatus of the molecular data processing method of the embodiment shown.

[0167] Figure 7 This is a schematic diagram of a molecular data processing apparatus according to an embodiment of the present disclosure, such as... Figure 4 As shown, the molecular data processing device 700 may include: a first acquisition unit 701, a second acquisition unit 702, a determination unit 703, and a generation unit 704.

[0168] The first acquisition unit 701 is used to collect data from the source domain and the target domain to obtain a molecular dataset.

[0169] The second acquisition unit 702 is used to determine a first original molecular feature and a second original molecular feature from the molecular dataset, wherein the first original molecular feature is used to characterize the molecular properties of the molecular data in the source domain, and the second original molecular feature is used to characterize the molecular properties of the molecular data in the target domain.

[0170] The determining unit 703 is used to determine model parameters based on the correlation between the first original molecular feature and the second original molecular feature, wherein the correlation is updated as the feature values ​​of the first original molecular feature and / or the feature values ​​of the second original molecular feature change.

[0171] The generation unit 704 is used to train based on model parameters to generate a molecular property prediction model, wherein the molecular property prediction model is used to predict the molecular properties of the molecular data to be identified.

[0172] Optionally, the determining unit 703 includes: an updating module, used to update the first original molecular feature to the first target molecular feature based on the association relationship, and to update the second original molecular feature to the second target molecular feature; and an acquisition module, used to acquire model parameters based on the first target molecular feature and the second target molecular feature.

[0173] Optionally, the molecular dataset in the source domain includes a source domain support dataset and a source domain query dataset, and the molecular dataset in the target domain includes a target domain support dataset and a target domain query dataset. The source domain support dataset and the target domain support dataset are used to train a molecular property prediction model, and the source domain query dataset and the target domain query dataset are used to test the molecular property prediction model. The device also includes a building unit for constructing relationships based on the source domain support dataset, the source domain query dataset, the target domain support dataset, and the target domain query dataset.

[0174] Optionally, the building unit includes: a building module for building relationships based on the source domain supporting dataset, a portion of the source domain query dataset in the source domain query dataset, the target domain supporting dataset, and a portion of the target domain query dataset in the target domain query dataset.

[0175] Optionally, the association relationship includes the similarity between molecular properties of molecular data in the same domain and / or the similarity between molecular properties of molecular data in different domains. The same domain includes the source domain or the target domain, and different domains include the source domain and the target domain. The construction module includes: a first acquisition submodule, used to acquire the union of the source domain support dataset, a portion of the source domain query dataset in the source domain query dataset, the target domain support dataset, and a portion of the target domain query dataset in the target domain query dataset; and a second acquisition submodule, used to acquire the original similarity between molecular properties of molecular data in the same domain and / or the original similarity between molecular properties of molecular data in different domains.

[0176] Optionally, the device further includes: a first processing unit, configured to perform sparsification processing on the original similarity between molecular properties of molecular data in the same domain of the union to obtain a target similarity between molecular data in the same domain of the union, wherein the target similarity is greater than a first similarity threshold; and / or, a second processing unit, configured to perform sparsification processing on the original similarity between molecular properties of molecular data in different domains of the union to obtain a target similarity between molecular data in different domains of the union, wherein the target similarity is greater than a second similarity threshold.

[0177] Optionally, the determining unit 703 includes: a first determining module, used to determine the loss function on the source domain supporting dataset, the loss function on the source domain query dataset, the loss function on the target domain supporting dataset, and the loss function on the target domain query dataset based on the association relationship; and a second determining module, used to determine the model parameters based on the loss function on the source domain supporting dataset, the loss function on the source domain query dataset, the loss function on the target domain supporting dataset, and the loss function on the target domain query dataset.

[0178] Optionally, the model parameters include first model parameters and second model parameters. The first model parameters are used to learn the characteristics of task pairs, and the second model parameters are used to learn the commonalities of multiple task pairs. The task pairs include molecular data in the source domain and molecular data in the target domain. The second determination module includes: a first determination submodule, used to determine the first model parameters based on the loss function on the source domain support dataset and the loss function on the target domain support dataset; and a second determination submodule, used to determine the second model parameters based on the loss function on the source domain query dataset, the loss function on the target domain query dataset, and the first model parameters.

[0179] Optionally, the first acquisition unit 701 includes: a sampling unit, used to randomly sample from the source domain to obtain molecular data in the source domain, and to randomly sample from the target domain to obtain molecular data in the target domain.

[0180] In the molecular data processing apparatus of this embodiment, a first acquisition unit is used to collect data from a source domain and a target domain to obtain a molecular dataset; a second acquisition unit is used to determine a first original molecular feature and a second original molecular feature from the molecular dataset, wherein the first original molecular feature is used to characterize the molecular properties of the molecular data in the source domain, and the second original molecular feature is used to characterize the molecular properties of the molecular data in the target domain; a determination unit is used to determine model parameters based on the correlation between the first original molecular feature and the second original molecular feature, wherein the correlation is updated as the feature values ​​of the first original molecular feature and / or the feature values ​​of the second original molecular feature change; and a generation unit is used to train the model based on the model parameters to generate a molecular property prediction model, wherein the molecular property prediction model is used to predict the molecular properties of the molecular data to be identified, thereby solving the technical problem of low accuracy of the model in predicting molecular properties and achieving the technical effect of improving the accuracy of the model in predicting molecular properties.

[0181] This disclosure also provides a molecular property prediction apparatus for performing the molecular property prediction method of the embodiment shown in FIG2.

[0182] Figure 8 This is a schematic diagram of a molecular property prediction device according to an embodiment of the present disclosure, such as... Figure 6As shown, the molecular property prediction device 800 may include a third acquisition unit 801 and a prediction unit 802.

[0183] The third acquisition unit 801 is used to acquire molecular data to be predicted.

[0184] The prediction unit 802 is used to input molecular data into the molecular property prediction model for prediction to obtain the molecular properties of the molecular data. The molecular property prediction model is determined based on the above-mentioned molecular data processing method.

[0185] In the molecular data processing apparatus of this embodiment, the third acquisition unit is used to acquire molecular data to be predicted; the prediction unit is used to input the molecular data into the molecular property prediction model for prediction to obtain the molecular properties of the molecular data, thereby solving the technical problem of low accuracy of the model in predicting molecular properties and achieving the technical effect of improving the accuracy of the model in predicting molecular properties.

[0186] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0187] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0188] Figure 9 This is a schematic diagram of an electronic device for a method of processing molecular data according to an embodiment of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0189] like Figure 9 As shown, device 900 includes a computing unit 901, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 902 or a computer program loaded into random access memory (RAM) 903 from storage unit 908. RAM 903 may also store various programs and data required for the operation of device 900. The computing unit 901, ROM 902, and RAM 903 are interconnected via bus 904. Input / output (I / O) interface 905 is also connected to bus 904.

[0190] Multiple components in device 900 are connected to I / O interface 905, including: input unit 906, such as keyboard, mouse, etc.; output unit 907, such as various types of monitors, speakers, etc.; storage unit 908, such as disk, optical disk, etc.; and communication unit 909, such as network card, modem, wireless transceiver, etc. Communication unit 909 allows device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0191] The computing unit 901 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 901 performs the various methods and processes described above, such as methods for processing molecular data. For example, in some embodiments, the methods for processing molecular data can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by the computing unit 901, one or more steps of the methods for processing molecular data described above can be performed. Alternatively, in other embodiments, the computing unit 901 can be configured to perform methods for processing molecular data by any other suitable means (e.g., by means of firmware).

[0192] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0193] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0194] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0195] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0196] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0197] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0198] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0199] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for processing molecular data, comprising: Molecular datasets are obtained by collecting data from the source and target domains; From the molecular dataset, a first original molecular feature contained in the source domain and a second original molecular feature contained in the target domain are determined, wherein the first original molecular feature is used to characterize the molecular properties of the molecular data in the source domain, and the second original molecular feature is used to characterize the molecular properties of the molecular data in the target domain; Based on the correlation between the first original molecular feature and the second original molecular feature, model parameters are determined, wherein the correlation is updated as the feature values ​​of the first original molecular feature and / or the feature values ​​of the second original molecular feature change; Based on the model parameters, a molecular property prediction model is generated through training, wherein the molecular property prediction model is used to predict the molecular properties of the molecular data to be identified. The molecular dataset in the source domain includes a source domain support dataset and a source domain query dataset, and the molecular dataset in the target domain includes a target domain support dataset and a target domain query dataset. The source domain support dataset and the target domain support dataset are used to train the molecular property prediction model, and the source domain query dataset and the target domain query dataset are used to test the molecular property prediction model. The method further includes: constructing the association relationship based on the source domain support dataset, the source domain query dataset, the target domain support dataset, and the target domain query dataset.

2. The method according to claim 1, wherein, Based on the correlation between the first original molecular feature and the second original molecular feature, the model parameters are determined as follows: Based on the aforementioned association, the first original molecular feature is updated to the first target molecular feature, and the second original molecular feature is updated to the second target molecular feature; The model parameters are obtained based on the first target molecule features and the second target molecule features.

3. The method according to claim 1, wherein, Based on the source domain supporting dataset, the source domain query dataset, the target domain supporting dataset, and the target domain query dataset, the association relationship is constructed, including: The association relationship is constructed based on the source domain supporting dataset, a portion of the source domain query dataset, the target domain supporting dataset, and a portion of the target domain query dataset.

4. The method according to claim 3, wherein, The association relationship includes the similarity between molecular properties of molecular data in the same domain and / or the similarity between molecular properties of molecular data in different domains. The same domain includes the source domain or the target domain, and the different domains include the source domain and the target domain. The association relationship is constructed based on the source domain support dataset, a portion of the source domain query dataset in the source domain query dataset, the target domain support dataset, and a portion of the target domain query dataset. Obtain the union of the source domain supporting dataset, a portion of the source domain query dataset in the source domain query dataset, the target domain supporting dataset, and a portion of the target domain query dataset in the target domain query dataset; Obtain and concentrate the original similarity between molecular properties of molecular data in the same domain and / or the original similarity between molecular properties of molecular data in different domains.

5. The method according to claim 4, wherein, The method further includes: The original similarity between molecular properties of molecular data in the same domain within the union set is sparsified to obtain the target similarity between the molecular data in the same domain within the union set, wherein the target similarity is greater than a first similarity threshold; and / or, The original similarity between molecular properties in the different domains of the union set is sparsified to obtain the target similarity between the molecular data in the different domains of the union set, wherein the target similarity is greater than a second similarity threshold.

6. The method according to claim 1, wherein, Based on the correlation between the first original molecular feature and the second original molecular feature, the model parameters are determined, including: Based on the aforementioned association, the loss function on the source domain supporting dataset, the loss function on the source domain query dataset, the loss function on the target domain supporting dataset, and the loss function on the target domain query dataset are determined respectively. The loss function on the source domain supporting dataset is calculated using the same formula as the loss function on the source domain supporting dataset. The loss function on the target domain supporting dataset is also obtained using the same formula as the loss function on the source domain query dataset. The model parameters are determined based on the loss function on the source domain supporting dataset, the loss function on the source domain query dataset, the loss function on the target domain supporting dataset, and the loss function on the target domain query dataset.

7. The method according to claim 6, wherein, The model parameters include first model parameters and second model parameters. The first model parameters are used to learn the characteristics of task pairs, and the second model parameters are used to learn the commonalities of multiple task pairs. The task pairs include molecular data in the source domain and molecular data in the target domain. The model parameters are determined based on the loss function on the source domain support dataset, the loss function on the source domain query dataset, the loss function on the target domain support dataset, and the loss function on the target domain query dataset, including: The first model parameters are determined based on the loss function on the source domain support dataset and the loss function on the target domain support dataset; The second model parameters are determined based on the loss function on the source domain query dataset, the loss function on the target domain query dataset, and the first model parameters.

8. The method according to any one of claims 1 to 7, wherein, The molecular dataset, obtained by collecting data from the source and target domains, includes: Molecular data in the source domain are obtained by random sampling from the source domain, and molecular data in the target domain are obtained by random sampling from the target domain.

9. A method for predicting molecular properties, comprising: Obtain molecular data to be predicted; The molecular data is input into a molecular property prediction model for prediction to obtain the molecular properties of the molecular data, wherein the molecular property prediction model is determined based on the method described in any one of claims 1 to 8.

10. A molecular data processing apparatus, comprising: The first acquisition unit is used to collect data from the source domain and the target domain to obtain a molecular dataset; The second acquisition unit determines a first original molecular feature and a second original molecular feature from the molecular dataset, wherein the first original molecular feature is used to characterize the molecular properties of the molecular data in the source domain, and the second original molecular feature is used to characterize the molecular properties of the molecular data in the target domain. The determining unit is used to determine model parameters based on the correlation between the first original molecular feature and the second original molecular feature, wherein the correlation is updated as the feature value of the first original molecular feature and / or the feature value of the second original molecular feature changes; A generation unit is used to train based on the model parameters to generate a molecular property prediction model, wherein the molecular property prediction model is used to predict the molecular properties of the molecular data to be identified. The molecular dataset in the source domain includes a source domain support dataset and a source domain query dataset, and the molecular dataset in the target domain includes a target domain support dataset and a target domain query dataset. The source domain support dataset and the target domain support dataset are used to train the molecular property prediction model, and the source domain query dataset and the target domain query dataset are used to test the molecular property prediction model. The device further includes a construction unit for constructing the association relationship based on the source domain support dataset, the source domain query dataset, the target domain support dataset, and the target domain query dataset.

11. The apparatus according to claim 10, wherein, The determining unit includes: An update module is used to update the first original molecular feature to a first target molecular feature based on the association relationship, and to update the second original molecular feature to a second target molecular feature. The acquisition module is used to acquire the model parameters based on the first target molecule features and the second target molecule features.

12. The apparatus according to claim 10, wherein, The building unit includes: The construction module is used to construct the association relationship based on the source domain supporting dataset, a portion of the source domain query dataset in the source domain query dataset, the target domain supporting dataset, and a portion of the target domain query dataset in the target domain query dataset.

13. The apparatus according to claim 12, wherein, The association relationship includes the similarity between molecular properties of molecular data in the same domain and / or the similarity between molecular properties of molecular data in different domains. The same domain includes the source domain or the target domain, and the different domains include the source domain and the target domain. The construction module includes: The first acquisition submodule is used to acquire the union of the source domain supporting dataset, a portion of the source domain query dataset in the source domain query dataset, the target domain supporting dataset, and a portion of the target domain query dataset in the target domain query dataset; The second acquisition submodule is used to acquire and collect the original similarity between molecular properties of molecular data in the same domain and / or the original similarity between molecular properties of molecular data in different domains.

14. The apparatus according to claim 13, wherein, The device further includes: A first processing unit is configured to perform sparsification processing on the original similarity between molecular properties of the molecular data in the same domain within the union set, to obtain a target similarity between the molecular data in the same domain within the union set, wherein the target similarity is greater than a first similarity threshold; and / or, The second processing unit is used to perform sparsification processing on the original similarity between molecular properties of molecular data in different domains in the union set to obtain the target similarity between molecular data in different domains in the union set, wherein the target similarity is greater than a second similarity threshold.

15. A device for predicting molecular properties, comprising: The third acquisition unit is used to acquire molecular data to be predicted; A prediction unit is used to input the molecular data into a molecular property prediction model for prediction to obtain the molecular properties of the molecular data, wherein the molecular property prediction model is determined based on the method described in any one of claims 1 to 8.

16. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-9.

17. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-9.

18. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-9.

Images