Method, device, equipment and storage medium for discovering lead compounds

Abstract

The application provides a lead compound discovery method, device, equipment and storage medium, comprising: obtaining a meta-learning model with initial parameters; obtaining activity data of a target target point and a plurality of historical target points, dividing the activity data of the first target point to obtain a training set and an original test set corresponding to the first target point, in the meta-learning method, for any one of the plurality of original test samples, selecting a training sample corresponding to the original test sample from the plurality of training samples, and mixing the original test sample and the training sample corresponding to the original test sample to obtain a mixed processing result; optimizing the initial parameters according to the mixed processing result of all first target points to obtain first parameters; and determining a lead compound for the target target point according to the first parameters, the meta-learning model and the activity data of the target target point, so that the determined lead compound has higher accuracy.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence (AI) technology, and in particular to methods, apparatus, devices and storage media for discovering lead compounds. Background Technology

[0002] Figure 1 A schematic diagram of the drug discovery process, such as... Figure 1 As shown, the typical drug discovery process includes: target identification and confirmation, lead compound discovery, lead compound discovery and optimization, candidate compound confirmation and development, and clinical trials. Currently, meta-learning methods can be used to learn a meta-learning model for lead compound discovery. This model characterizes the relationship between the molecular structure of the lead compound and its activity against the target.

[0003] Because there is currently too little test data for training meta-learning models, and the gap between training and test data is large, meta-learning models tend to memorize test data rather than extract the relationship between molecular structure and activity value from training data. This results in low performance of the obtained meta-learning models, which in turn leads to low accuracy of the lead compounds identified based on these models. Summary of the Invention

[0004] This application provides a method, apparatus, device, and storage medium for discovering lead compounds, which results in higher accuracy of the discovered lead compounds.

[0005] In a first aspect, a method for discovering lead compounds is provided, comprising: acquiring a meta-learning model with initial parameters; acquiring activity data for a target target and multiple historical targets, wherein for any one of the target target and multiple historical targets, each activity data for the target includes: the molecular structure of the lead compound for the target and the activity value of the molecular structure for the target; dividing the activity data of the first target to obtain a training set and an original test set corresponding to the historical targets, wherein the training set includes multiple training samples and the original test set includes multiple original test samples, wherein the first target is any one of the multiple historical targets, or the first target is the target target and any one of the multiple historical targets; in the meta-learning method, for any one of the multiple original test samples, selecting the training sample corresponding to the original test sample from the multiple training samples, and performing a mixed processing on the original test sample and the training sample corresponding to the original test sample to obtain a mixed processing result; optimizing the initial parameters according to the mixed processing results of all the first targets to obtain first parameters; and determining the lead compound for the target target based on the first parameters, the meta-learning model, and the activity data of the target target.

[0006] Secondly, a device for discovering lead compounds is provided, comprising: a first acquisition module, a second acquisition module, a partitioning module, a mixing module, an optimization module, and a determination module. The first acquisition module is used to acquire a meta-learning model with initial parameters. The second acquisition module is used to acquire activity data for a target and multiple historical targets. For any one of the target and historical targets, each activity data for the target includes: the molecular structure of the lead compound for the target and the activity value of the molecular structure for the target. The partitioning module is used to partition the activity data of the first target to obtain a training set and an original test set corresponding to the first target. The training set includes multiple training samples, and the original test set includes multiple training samples. The original test sample is used, where the first target is any one of multiple historical targets, or the first target is the target target and any one of multiple historical targets. The hybrid processing module is used in the meta-learning method to select the training sample corresponding to the original test sample from multiple training samples for any one of the original test samples, and to perform hybrid processing on the original test sample and the training sample corresponding to the original test sample to obtain the hybrid processing result. The optimization module is used to optimize the initial parameters according to the hybrid processing results of each of the first targets to obtain the first parameters. The determination module is used to determine the lead compound for the target target according to the first parameters, the meta-learning model and the activity data of the target target.

[0007] Thirdly, an electronic device is provided, comprising: a processor and a memory for storing a computer program, the processor for calling and running the computer program stored in the memory, and performing methods as described in various implementations of the first aspect.

[0008] Fourthly, a computer-readable storage medium is provided for storing a computer program that causes a computer to perform the methods described in the first aspect or its various implementations.

[0009] Fifthly, a computer program product is provided, including computer program instructions that cause a computer to perform the methods as described in the first aspect or its various implementations.

[0010] Sixthly, a computer program is provided that causes a computer to perform the methods described in the first aspect or its various implementations.

[0011] By using the technical solution of this application, the mixing of the test set and the training set reduces the memorization of the test set by the meta-learning model, improves the generalization of the model, and enables the model to learn the relationship between molecular structure and activity value from the training set, thereby improving the performance of the meta-learning model. Consequently, the accuracy of the lead compounds identified based on this meta-learning model is higher.

[0012] Furthermore, if the electronic device performs random mixing of the data from the original test set and the training set, such as random linear mixing, this randomness can further prevent the meta-learning model from memorizing the test set, allowing the model to learn the relationship between molecular structure and activity value from the training set, thereby improving the performance of the meta-learning model. Consequently, the accuracy of the lead compounds identified based on this meta-learning model is higher. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the drug discovery process;

[0015] Figure 2 A flowchart illustrating a method for discovering a lead compound, provided as an embodiment of this application;

[0016] Figure 3 A flowchart of the hybrid processing procedure provided in the embodiments of this application;

[0017] Figure 4 A flowchart of an initial parameter optimization method provided in an embodiment of this application;

[0018] Figure 5 A flowchart of an initial parameter optimization method provided in another embodiment of this application;

[0019] Figure 6 A schematic diagram of a lead compound detection device 600 provided for an embodiment of this application;

[0020] Figure 7 This is a schematic block diagram of the electronic device 700 provided in the embodiments of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0023] It should be understood that the technical solution of this application relates to the field of AI technology.

[0024] AI (Artificial Intelligence) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, artificial intelligence is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. Artificial intelligence studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0025] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0026] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0027] The drug discovery process typically includes: target identification and confirmation, lead compound discovery, lead compound discovery and optimization, candidate compound confirmation and development, and clinical trials. Sometimes, an animal testing phase is added between the lead compound discovery and optimization phase and the candidate compound confirmation and development phase.

[0028] The drug discovery process begins with identifying a hypothesis about a disease, typically by discovering a dysfunctional protein that causes the disease or pathology. Proteins are the backbone of the human body, performing most of the biological tasks necessary for survival: from synthesizing molecules and mediating signals between cells and tissues to fighting infection. Most diseases are caused by the malfunction of one or more proteins to some extent: they may have the wrong shape or chemical composition, leading to incorrect biochemical reactions and resulting in systemic damage. Even a slight change in a protein can have serious consequences, even death. Proteins that play a role in the development and progression of disease are known as targets.

[0029] Once a target is identified, intensive follow-up studies must be conducted to prove the selection was correct; this process is called target identification and validation, or target verification. This work includes a variety of studies, from determining the crystal structure of the target protein to confirming its association with related diseases. The link between the target and the disease is a crucial step in new drug development, and it can determine the success or failure of the entire program.

[0030] After target identification and confirmation, the next step is to discover methods that affect the faulty protein, typically by blocking or altering its activity. This involves large-scale screening programs where thousands or millions of compounds are tested to see if they can affect the target in a beneficial way. Molecules with acceptable activity are called lead compounds of the drug target. Of these lead compounds, most will prove to be false positives, and only a small fraction will ultimately be confirmed and screened as lead compounds.

[0031] Although the lead compounds have shown significant activity at the target sites, their other key parameters still need to be optimized, such as metabolic stability, safety, bioavailability, and other properties.

[0032] The most crucial step in lead drug optimization is developing one or a group of molecules prepared for preclinical studies. These molecules are then tested on animals to see if they can sustain their intended effects in real organisms. If the activity and safety of the lead drug are confirmed in animal studies, this molecule is ultimately nominated as a drug candidate. It can then proceed to preclinical studies, the final step in the preclinical drug discovery process.

[0033] As mentioned above, the second step in drug discovery is the discovery of lead compounds targeting target proteins. Compared to traditional experimental screening of lead compounds, virtual screening eliminates the need for compound samples, significantly saving manpower and resources and accelerating the screening process. Ligand-based drug design is a common method in virtual screening, which involves learning and building a model of the relationship between molecular structure and activity from known active ligand small molecule structures to predict lead compounds. However, the limited amount of measured activity data for targets severely restricts the accuracy of the model. Considering the existence of molecular structures with measured activity against other target proteins—that is, other target proteins with activity data—including the molecular structure of lead compounds targeting that target protein and the activity value of that molecular structure against that target, we consider introducing meta-learning algorithms into ligand-based drug design. However, due to the limited amount of test data currently available for training meta-learning models, and the significant gap between training and test data, meta-learning models tend to memorize test data rather than extract the relationship between molecular structure and activity values ​​from training data. This results in lower performance of the obtained meta-learning models, leading to lower accuracy in identifying lead compounds based on these models.

[0034] To overcome these problems, this application proposes a method in which the test set and training set of the meta-learning model are mixed to optimize the model. This mixed processing reduces the meta-learning model's reliance on the test set, improves its generalization ability, and allows the model to learn the relationship between molecular structure and activity values ​​from the training set. This enhances the performance of the meta-learning model, leading to higher accuracy in identifying lead compounds based on it.

[0035] The technical solution of this application will be described in detail below:

[0036] Figure 2 This flowchart illustrates a method for discovering lead compounds, as provided in an embodiment of this application. The subject of this method can be any electronic device, including various types of terminal devices or servers. Terminal devices can be smartphones, tablets, laptops, desktop computers, smart speakers, wearable devices, etc., but are not limited to these. Servers can be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers providing cloud computing services, but are not limited to these. Figure 2 As shown, the method includes the following steps:

[0037] S210: Obtain the meta-learning model with initial parameters;

[0038] S220: Obtain activity data for the target and multiple historical targets. For any one of the target and multiple historical targets, the activity data for each target includes: the molecular structure of the lead compound for the target and the activity value of the molecular structure for the target.

[0039] S230: Divide the activity data of the first target to obtain the training set and the original test set corresponding to the first target. The training set includes multiple training samples, and the original test set includes multiple original test samples. The first target is any one of the above-mentioned historical targets, or the first target is the target target and any one of the multiple historical targets.

[0040] S240: In the meta-learning method, for any one of the multiple original test samples, the training sample corresponding to the original test sample is selected from the multiple training samples, and the original test sample and the training sample corresponding to the original test sample are mixed to obtain the mixed processing result.

[0041] S250: The initial parameters are optimized based on the mixed processing results of all first target points to obtain the first parameters.

[0042] S260: Identify lead compounds for the target based on the first parameter, the meta-learning model, and the activity data of the target.

[0043] Optionally, the molecular structure of a lead compound can be represented by the characteristics of that molecular structure; that is, in this application, the molecular structure of a lead compound can be understood as the characteristics of that molecular structure.

[0044] Optionally, there are usually multiple features of the molecular structure. This application may use 1024-dimensional Morgan fingerprints developed by Rdkit as features of the molecular structure, but is not limited to this.

[0045] It should be understood that for any one of the aforementioned historical targets, the activity value (PIC50) of the lead compound against that historical target has been determined. A higher activity value indicates a stronger inhibitory effect of the lead compound against that historical target, and thus a better overall effect.

[0046] It should be understood that the target site refers to the target site for which the lead compound needs to be identified.

[0047] Optionally, the first target has multiple activity data, and the electronic device can divide the multiple activity data of the first target into a training set and an original test set according to a certain ratio, such as an 8:2 ratio.

[0048] It should be understood that this application does not impose any restrictions on the ratio of data in the training set to the original test set.

[0049] It should be understood that, according to the definition of activity data, any training sample in the training set includes: the molecular structure of a lead compound targeting the first target and the activity value of that molecular structure targeting the first target. Similarly, any original test sample in the original test set also includes: the molecular structure of a lead compound targeting the first target and the activity value of that molecular structure targeting the first target.

[0050] It should be noted that the data included in each training sample or original test sample are usually not exactly the same. For example, training sample A includes the molecular structure 1 of the lead compound 1 targeting historical target 1 and the activity value of the molecular structure 1 against historical target 1. Training sample B includes the molecular structure 2 of the lead compound 2 targeting historical target 1 and the activity value of the molecular structure 2 against historical target 1.

[0051] Optionally, meta-learning methods can be, but are not limited to, the following:

[0052] The purpose of this application is to fully utilize the activity data of all previous historical targets. Therefore, it is assumed that all historical targets and the target target use the same meta-learning model, the difference being that each target has its own parameters. For example, the parameters of the meta-learning model corresponding to the first historical target are... The parameters of the meta-learning model corresponding to the second historical target are: And so on, the parameters of the meta-learning model corresponding to the nth historical target are... And so on. In order to fully utilize the knowledge of historical targets, it is assumed that the parameters of the meta-learning model corresponding to all targets all originate from a common initial parameter. The parameters of the meta-learning model corresponding to the historical target point are obtained through single-step or multi-step gradient optimization. For example, for the k-th historical target point, the electronic device uses the training data corresponding to that historical target point and employs single-step gradient optimization to obtain the parameters. See the following formula for details, but it is not limited to this:

[0053] (1)

[0054] in, These are the parameters of the meta-learning model corresponding to that historical target, i.e., the second parameter in the following text. The training set for this historical target is a gradient-optimized learning rate. , This represents the set of molecular structures of the lead compounds included in each activity data point in the training set. This indicates the activity value included in each active data point in the training set.

[0055] Suppose that for any one of multiple historical targets, there are N activity data points. , ,in, This indicates the molecular structure of the lead compound corresponding to the historical target. As mentioned above, this molecular structure is equivalent to the features of the molecular structure. There are usually multiple features of the molecular structure. This application may use the 1024-dimensional Morgan fingerprint developed by Rdkit as the features of the molecular structure, but is not limited to this. This represents the activity value (PIC50) of the molecular structure against the historical target. A higher activity value indicates a stronger inhibitory effect of the lead compound against the historical target, and thus a better outcome. Furthermore, the electronic device can train a meta-learning model based on training data from this historical target. The objective function for training, i.e., the loss function, is as follows, but is not limited to:

[0056] (2)

[0057] It should be noted that since this loss function refers to the loss function of the meta-learning model for any historical target point, the parameters of the meta-learning model... This is also a general term. So, let's assume the parameters for the corresponding meta-learning model are... The historical target, its training set includes N training samples, then according to the above formula (2), we know that

[0058] (3)

[0059] It should be understood that, as mentioned above, the electronic device can obtain the parameters of the meta-learning model corresponding to the historical target point through the training data corresponding to the historical target point using a multi-step gradient optimization method. The multi-step gradient optimization process is as follows: First, the electronic device performs single-step gradient optimization using formula (1) to obtain the parameters. Secondly, the parameters in formula (1) Replace with parameters To obtain new parameters This completes the second step of gradient optimization, and so on, until the final parameters are obtained.

[0060] Furthermore, in order to optimize the meta-learning models corresponding to all targets through a very limited number of gradient optimization steps starting from the initial parameters, the meta-learning method uses the performance of the meta-learning model corresponding to each historical target (the parameters of the model are the parameters of the meta-learning model corresponding to the historical target) on its test data as feedback to optimize the initial parameters. Specifically, it can be optimized through the following formula (4) or formula (5), but is not limited to these:

[0061] (4)

[0062] in, This refers to the optimized parameters, i.e., the first parameter in the following text. Indicates the initial parameters. Indicates the number of historical targets. This represents the original test set for the k-th historical target. This represents the set of molecular structures of the lead compounds included in each activity data point in the original test set. This indicates the activity values ​​included in each activity data point in the original test set. This refers to a preset factor, which can be set according to actual conditions; this application does not impose any restrictions on it. This represents the parameters of the meta-learning model corresponding to the historical target, i.e., the second parameter in the following text, regarding... The method for finding can be found in formula (3), which will not be elaborated further in this application.

[0063] (5)

[0064] in, This refers to the optimized parameters, i.e., the first parameter in the following text. Indicates the initial parameters. Indicates the number of historical targets. This represents the original test set for the k-th historical target. This represents the set of molecular structures of the lead compounds included in each activity data point in the original test set. This indicates the activity values ​​included in each activity data point in the original test set. This represents the training set for the k-th historical target. This represents the set of molecular structures of the lead compounds included in each activity data point in the training set. This represents the activity value included in each active data point in the training set. This refers to a preset factor, which can be set according to actual conditions; this application does not impose any restrictions on it. This represents the gradient optimization learning rate. This represents the parameters of the meta-learning model corresponding to the historical target, i.e., the second parameter in the following text, regarding... The method for finding can be found in formula (3), which will not be elaborated further in this application.

[0065] Furthermore, targeting specific targets, their activity data can be used to analyze... By performing finite-step gradient optimization, the parameters of the target prediction model corresponding to the target point are finally obtained, namely the third parameter. The target prediction model with the third parameter is the target prediction model corresponding to the target point. The lead compound of the target point can be determined through this model.

[0066] However, due to the limited amount of test data currently available for training meta-learning models, and the significant gap between training and test data, meta-learning models tend to memorize test data rather than extract the relationship between molecular structure and activity values ​​from training data. This results in lower performance of the obtained meta-learning models, leading to lower accuracy in identifying lead compounds based on these models.

[0067] To overcome these problems, this application proposes a method in which the test set and training set of the meta-learning model are mixed to optimize the meta-learning model. The mixing process is described in detail below:

[0068] Figure 3 A flowchart of the hybrid processing procedure provided in the embodiments of this application is shown below. Figure 3 As shown, the process includes the following steps:

[0069] S310: Perform gradient optimization on the initial parameters of the meta-learning model based on the training set corresponding to the first target point to obtain the second parameters;

[0070] S320: For any one of the multiple original test samples, select the training sample corresponding to the original test sample from the multiple training samples, and input the first molecular structure and the second molecular structure into the first neural network model respectively to obtain the first output results of the first molecular structure and the second molecular structure respectively.

[0071] S330: Mix the first output result of the first molecular structure and the first output result of the second molecular structure to obtain the first mixed processing result;

[0072] S340: Input the first hybrid processing result into the second neural network model to obtain the second output result;

[0073] S350: The activity values ​​of the original test samples and the activity values ​​of the corresponding training samples are mixed to obtain a second mixed processing result.

[0074] The above-mentioned mixing process of the original test sample and the training sample corresponding to the original test sample results in the following mixed processing results: the second output result and the second mixed processing result.

[0075] It should be noted that, as mentioned above, the first target can be any one of multiple historical targets, or the first target can be either the target target or any one of multiple historical targets. If the first target is any one of multiple historical targets, the explanation of S310 can refer to the process in the meta-learning method described above, where the electronic device obtains the parameters of the meta-learning model corresponding to the historical target through the training data corresponding to the historical target and uses a single-step or multi-step gradient optimization method. This application will not elaborate on this further. If the first target is either the target target or any one of multiple historical targets, the explanation of S310 can also refer to the process in the meta-learning method described above, where the electronic device obtains the parameters of the meta-learning model corresponding to the historical target through the training data corresponding to the historical target and uses a single-step or multi-step gradient optimization method. The difference is that formula (1) also applies to the target target.

[0076] It should be understood that the first molecular structure is the molecular structure of the original test sample, and the second molecular structure is the molecular structure of the training sample corresponding to the original test sample. The first neural network model consists of the first layer to the target layer of the meta-learning model with the second parameter; that is, the first neural network model is a sub-model of the meta-learning model. The second neural network model consists of the layer after the target layer to the last layer of the meta-learning model with the second parameter; that is, the second neural network model is a sub-model of the meta-learning model.

[0077] Optionally, the electronic device may, for any given original test sample, select the training sample corresponding to the original test sample from among multiple training samples in any of the following feasible manners, but is not limited thereto:

[0078] Option 1: The electronic device can randomly select the corresponding training sample from multiple training samples for any given original test sample. For example, the training sample corresponding to the original test sample i selected by the electronic device using a random selection method is training sample j.

[0079] Option 2: The electronic device can sort the original test samples in the original test set and also sort the training samples in the training set, using the training sample with the same number as the original test sample as the corresponding training sample. For example: Suppose the electronic device sorts N original test samples in the original test set, where the N original test samples are: original test sample 1, original test sample 2, ..., original test sample N. The electronic device also sorts N training samples in the training set, where the N training samples are: training sample 1, training sample 2, ..., training sample N. Based on this, the training sample corresponding to original test sample 1 is training sample 1, the training sample corresponding to original test sample 2 is training sample 2, and so on, until the training sample corresponding to original test sample N is training sample N.

[0080] Optionally, the sorting rules for the original test samples and the sorting rules for the training samples by the electronic device may be the same or different, and this application does not impose any restrictions on this. Furthermore, this application does not impose any restrictions on the sorting rules for the original test samples and the sorting rules for the training samples.

[0081] Option 3: The electronic device can sort the original test samples in the original test set and also sort the training samples in the training set. The training sample corresponding to the i-th original test sample is determined as the Ni-th training sample, where i is a positive integer and N is the number of original test samples and training samples sorted. For example, suppose the electronic device sorts N original test samples in the original test set, where the N original test samples are: original test sample 1, original test sample 2, ..., original test sample N. The electronic device also sorts N training samples in the training set, where the N training samples are: training sample 1, training sample 2, ..., training sample N. Based on this, the training sample corresponding to original test sample 1 is training sample N, the training sample corresponding to original test sample 2 is training sample N-1, and so on, until the training sample corresponding to original test sample N is training sample 1.

[0082] It should be understood that the original test samples and these multiple training samples mentioned above are samples from the original test set and training set corresponding to the same target, respectively.

[0083] It should be understood that since the meta-learning model is a multi-layer deep neural network model, such as a three-layer deep neural network model, the two middle hidden layers can each contain 500 neurons. Therefore, in the meta-learning method, when an electronic device performs mixed processing on any original test sample and the corresponding training sample, it first needs to select the target layer corresponding to the original test sample.

[0084] Optionally, the electronic device may select the target layer in the following ways, but is not limited to:

[0085] One possible approach: The electronic device randomly selects the target layer in a multi-layer deep neural network model.

[0086] Option 2: The electronic device sorts the original test samples. According to the order of the original test samples, the first layer of the multilayer deep neural network model is determined as the target layer corresponding to the first original test sample, the second layer of the multilayer deep neural network model is determined as the target layer corresponding to the second original test sample, and so on. Assuming that the multilayer deep neural network model has M layers, then the Mth layer of the multilayer deep neural network model is determined as the target layer corresponding to the Mth original test sample. If the number of original test samples is N, and N is greater than M, then the first layer of the multilayer deep neural network model is determined as the target layer corresponding to the (M+1)th original test sample, the second layer of the multilayer deep neural network model is determined as the target layer corresponding to the (M+2)th original test sample, and so on, to determine the target layer corresponding to each original test sample.

[0087] Option 3: Assuming the multilayer deep neural network model has M layers, the electronic device sorts the original test samples. According to the order of the original test samples, the Mth layer in the multilayer deep neural network model is determined as the target layer corresponding to the first original test sample, the (M-1)th layer in the multilayer deep neural network model is determined as the target layer corresponding to the second original test sample, and so on. The first layer in the multilayer deep neural network model is determined as the target layer corresponding to the Mth original test sample. If the number of original test samples is N, and N is greater than M, then the Mth layer in the multilayer deep neural network model is determined as the target layer corresponding to the (M+1)th original test sample, the (M-1)th layer in the multilayer deep neural network model is determined as the target layer corresponding to the (M+2)th original test sample, and so on, to determine the target layer corresponding to each original test sample.

[0088] Optionally, the electronic device can perform linear mixing processing on the first output result of the first molecular structure and the first output result of the second molecular structure to obtain a first mixed processing result. Of course, the electronic device can also perform other mixing processing on the first output result of the first molecular structure and the first output result of the second molecular structure to obtain a first mixed processing result, and this application does not limit this.

[0089] Optionally, the electronic device can linearly mix the activity value of the original test sample with the activity value of the corresponding training sample to obtain a second mixing result. Of course, the electronic device can also perform other mixing processes on the activity value of the original test sample and the activity value of the corresponding training sample to obtain a second mixing result, and this application does not limit this.

[0090] Optionally, the linear mixing weight used by the electronic device to perform linear mixing processing on the first output result of the first molecular structure and the first output result of the second molecular structure can be the same as the linear mixing weight used to perform linear mixing processing on the activity value of the original test sample and the activity value of the training sample corresponding to the original test sample. Of course, it can also be different, and this application does not limit it.

[0091] It should be noted that the mixed processing method used by the electronic device for the molecular structure of the same pair of original test samples and the corresponding training samples can be the same as or different from the mixed processing method used for their activity values. This application does not impose any restrictions on this.

[0092] Optionally, if the electronic device performs a linear mixture processing on the original test samples and the corresponding training samples, then the coefficients of the linear mixture processing can be obtained by sampling from the beta distribution, i.e. , Represents the original test sample The corresponding linear mixing coefficients.

[0093] It should be noted that this application does not impose any restrictions on how the linear mixing coefficients of the original test samples are determined.

[0094] The following section uses the linear mixing of the first output results of the first molecular structure and the first output results of the second molecular structure by an electronic device as an example to illustrate the linear mixing process of the activity value of the original test sample and the activity value of the corresponding training sample:

[0095] For example, suppose the original test set for the first target is The training set for the first target is ,in, This represents the set of molecular structures of the lead compounds included in each activity data point in the original test set. This indicates the activity value included in each activity data point in the original test set; This represents the set of molecular structures of the lead compounds included in each activity data point in the training set. This represents the activity value included in each activity data point in the training set. Assume the original test set... Any of the original test samples in the sample is ,in, This indicates the molecular structure of the seed compounds included in the original test sample. This indicates the activity values ​​included in the original test sample. (Regarding this original test sample...) The training samples randomly selected by the electronic device from multiple training samples are That is, the original test sample The corresponding training samples are Assume the meta-learning model is an L-layer deep neural network, and the electronic device is the original test sample. The target layer is selected randomly. Layer, assuming the meta-learning model corresponding to the first target is It should be noted that for the multiple first targets and target targets mentioned above, their corresponding meta-learning model architectures are the same. The difference lies in the parameters of the meta-learning models corresponding to these targets. For example, in this example, the parameters corresponding to the first target are... Furthermore, electronic devices can... Layers And its corresponding training samples are Perform linear mixing processing as follows:

[0096] (6)

[0097] (7)

[0098] Among them, here It has no practical significance. This represents the first output result of the first molecular structure. This represents the first output result indicating the second molecular structure. This indicates the result of the first mixing process. This indicates the second output result. Represents the original test sample The activity value and its corresponding training sample The activity value is the result of linear mixing, which is the result of the second mixing treatment mentioned above.

[0099] The above-mentioned S250 can be implemented in any of the following ways, but is not limited to:

[0100] Method 1 can be implemented. Figure 4 A flowchart of an initial parameter optimization method provided in an embodiment of this application is shown below. Figure 4 As shown, the method includes the following steps:

[0101] S410: Determine the first loss function of the meta-learning model corresponding to the first target based on the second output result and the second hybrid processing result corresponding to the first target.

[0102] S420: Optimize the initial parameters according to the first loss function corresponding to each of the first target points to obtain the first parameters.

[0103] It should be understood that for the first target, there are multiple original test samples and multiple training samples. Each original test sample can be solved according to the above content to obtain the second output result and the second mixed processing result of the original test sample. Therefore, the second output result corresponding to the first target is usually multiple, and the second mixed processing result corresponding to the first target is also usually multiple.

[0104] Optionally, the electronic device may refer to the above formula (3) to solve the first loss function based on the second output result and the second mixed processing result corresponding to the first target point, but is not limited to this.

[0105] Optionally, the electronic device can calculate the first gradient value of the first loss function corresponding to each of the first target points with respect to the initial parameters; and optimize the initial parameters based on the first gradient value of the first loss function corresponding to each of the first target points with respect to the initial parameters to obtain the first parameters. The first gradient value can be a first-order gradient, but is not limited to this.

[0106] It should be understood that this application does not impose any restrictions on how the initial parameters are optimized based on the first loss function corresponding to each of the first targets to obtain the first parameters.

[0107] Optionally, the electronic device can sum the first gradient values ​​of the first loss function corresponding to each of the first target points with respect to the initial parameters to obtain a first summation result; and optimize the initial parameters based on the first summation result to obtain the first parameters. Alternatively, the electronic device can perform a weighted average of the first gradient values ​​of the first loss function corresponding to each of the first target points with respect to the initial parameters to obtain a weighted average value; and optimize the initial parameters based on the weighted average value to obtain the first parameters. In summary, this application does not limit how the initial parameters are optimized based on the first gradient values ​​of the first loss function corresponding to each of the first target points with respect to the initial parameters to obtain the first parameters.

[0108] Optionally, the electronic device can calculate the product of the first summation result and a preset factor to obtain the first product result; and optimize the initial parameters based on the first product result to obtain the first parameter. Alternatively, the electronic device can directly optimize the initial parameters based on the first summation result to obtain the first parameter. In summary, this application does not limit how the initial parameters are optimized based on the first summation result to obtain the first parameter.

[0109] Optionally, the electronic device can calculate the difference between the initial parameter and the first product result to obtain the first parameter. Alternatively, the electronic device can multiply the first product result by a coefficient, such as a coefficient greater than 0 and less than 1, to obtain a product result. The electronic device then calculates the difference between the initial parameter and this product result to obtain the first parameter. In short, this application does not limit how the initial parameter is optimized based on the first product result to obtain the first parameter.

[0110] For example, the electronic device can optimize its initial parameters using the following formula (8):

[0111] (8)

[0112] in, Indicates the first parameter. Indicates the initial parameters. Indicates the number of historical targets. This represents the original test set for the k-th historical target. This represents the first loss function of the meta-learning model corresponding to the first target point. This represents the first gradient value of the first loss function with respect to the initial parameters. This represents the first summation result. This refers to a preset factor, which can be set according to actual conditions; this application does not impose any restrictions on it. This represents the result of the first product. The second parameter represents the parameters of the meta-learning model corresponding to the first target point. The method for finding can be found in formula (3), which will not be elaborated further in this application.

[0113] Method 2 can be implemented. Figure 5 A flowchart of an initial parameter optimization method provided in another embodiment of this application is shown below. Figure 5 As shown, the method includes the following steps:

[0114] S510: Determine the first loss function of the meta-learning model corresponding to the first target based on the second output result and the second hybrid processing result corresponding to the first target.

[0115] S520: Determine the second loss function corresponding to the first target based on the training set corresponding to the first target;

[0116] S530: Optimize the initial parameters based on the first loss function and the second loss function corresponding to each of the first target points to obtain the first parameters.

[0117] It should be understood that for the first target, there are multiple original test samples and multiple training samples. Each original test sample can be solved according to the above content to obtain the second output result and the second mixed processing result of the original test sample. Therefore, the second output result corresponding to the first target is usually multiple, and the second mixed processing result corresponding to the first target is also usually multiple.

[0118] Optionally, the electronic device may refer to the above formula (3) to solve the first loss function based on the second output result and the second mixed processing result corresponding to the first target point, but is not limited to this.

[0119] Optionally, the electronic device may determine the second loss function corresponding to the first target point according to the training set corresponding to the first target point in accordance with the above formula (3), but is not limited to this.

[0120] Optionally, the electronic device can calculate the second gradient value of the first loss function corresponding to each of the first target points with respect to the first parameters; calculate the third gradient value of the second loss function corresponding to each of the first target points with respect to the initial parameters; and optimize the initial parameters based on the second and third gradient values ​​corresponding to each of the first target points to obtain the first parameters.

[0121] Optionally, the second gradient value mentioned above can be a first-order gradient, and the third gradient value can be a second-order single-step gradient result or a second-order multi-step gradient result. This application does not impose any restrictions on this.

[0122] Optionally, the electronic device can calculate the product of the third gradient value corresponding to the first target point and the gradient optimization learning rate to obtain the second product result; and optimize the initial parameters based on the second gradient value and the second product result corresponding to each of the first target points to obtain the first parameters. Alternatively, the electronic device can directly optimize the initial parameters based on the second and third gradient values ​​corresponding to each of the first target points to obtain the first parameters, and this application does not impose any restrictions on this.

[0123] Optionally, the electronic device can calculate the difference between the identity matrix and the second product result corresponding to the first target point to obtain a first difference result; and optimize the initial parameters based on the second gradient values ​​corresponding to each of the first target points and the first difference result to obtain the first parameters. Alternatively, the electronic device can multiply the second product result corresponding to the first target point by a coefficient, such as a coefficient greater than 0 and less than 1, to obtain a product result. The electronic device can then calculate the difference between the identity matrix and this product result to obtain a difference result, and optimize the initial parameters based on the second gradient values ​​corresponding to each of the first target points and the difference result to obtain the first parameters. In short, this application does not limit how the initial parameters are optimized based on the second gradient values ​​and the second product result corresponding to each of the first target points to obtain the first parameters.

[0124] Optionally, the electronic device can calculate the product of the second gradient value and the first difference result corresponding to the first target point to obtain a third product result; and optimize the initial parameters based on the third product results corresponding to each of the first target points to obtain the first parameters. Alternatively, the electronic device can calculate the product of the second gradient value and the first difference result corresponding to the first target point to obtain a third product result, multiply the third product result by a coefficient to obtain a product result; and optimize the initial parameters based on the product result corresponding to each of the first target points to obtain the first parameters. In summary, this application does not limit how the initial parameters are optimized based on the second gradient value and the first difference result corresponding to each of the first target points to obtain the first parameters.

[0125] Optionally, the electronic device can sum the third product results corresponding to each of the first target points to obtain a second summation result; and optimize the initial parameters based on the second summation result to obtain the first parameters. Alternatively, the electronic device can perform a weighted average of the third product results corresponding to each of the first target points to obtain a second summation result; and optimize the initial parameters based on the second summation result to obtain the first parameters. In summary, this application does not limit how the initial parameters are optimized based on the third product results corresponding to each of the first target points to obtain the first parameters.

[0126] Optionally, the electronic device calculates the product of the second summation result and a preset factor to obtain a fourth product result; the initial parameters are then optimized based on the fourth product result to obtain the first parameter. Alternatively, the electronic device can directly optimize the initial parameters based on the second summation result to obtain the first parameter. This application does not limit how the initial parameters are optimized based on the second summation result to obtain the first parameter.

[0127] Optionally, the electronic device can calculate the difference between the initial parameter and the fourth product result to obtain the first parameter. Alternatively, the electronic device can multiply the fourth product result by a coefficient, such as a coefficient greater than 0 and less than 1, to obtain a product result. The electronic device then calculates the difference between the initial parameter and this product result to obtain the first parameter. In short, this application does not limit how the initial parameter is optimized based on the fourth product result to obtain the first parameter.

[0128] For example, the electronic device can optimize the initial parameters using the following formula (9) to obtain the first parameter:

[0129] (9)

[0130] in, This refers to the optimized parameters, i.e., the first parameter. Indicates the initial parameters. Indicates the number of historical targets. This represents the training set for the k-th historical target. This represents the set of molecular structures of the lead compounds included in each activity data point in the training set. This represents the activity value included in each active data point in the training set. This represents the first loss function of the meta-learning model corresponding to the first target point. This represents the second loss function. This represents the second gradient value of the first loss function with respect to the first parameter. This represents the third gradient value of the second loss function with respect to the initial parameters. This represents the gradient optimization learning rate. This represents the result of the second product. Represents the identity matrix. This represents the first difference result. This represents the result of the third product. This represents the result of the second summation. This represents the result of the fourth product.

[0131] It should be understood that the variations of the above formulas (8) and (9) can also be used to determine the first parameter, and this application does not limit this.

[0132] Optionally, the electronic device can optimize the initial parameters once or multiple times. If multiple optimizations are performed, after each determination of the first parameter, the electronic device determines whether the parameter optimization stopping condition is met. If the parameter optimization stopping condition is met, the optimization process is stopped. If the parameter optimization stopping condition is not met, S210 is executed again, and the first parameter is used as the new initial parameter of the meta-learning model until the parameter optimization stopping condition is met.

[0133] Optionally, the above parameter optimization stopping condition may be that the current parameter optimization count has reached the specified number of parameter optimizations, or that the first parameter obtained each time constitutes a parameter sequence that converges. However, the parameter optimization stopping condition is not limited to these.

[0134] Optionally, after the electronic device obtains the first parameter, it can substitute the first parameter and the activity data corresponding to the target point, such as the training data, into formula (10), that is, use the single-step gradient optimization method to obtain the parameters of the meta-learning model corresponding to the target point, which is called the third parameter here.

[0135] (10)

[0136] in, Indicates the first parameter. Indicates the third parameter. The training set representing the target points. This represents the set of molecular structures of the lead compounds included in each activity data point in the training set. This indicates the activity value included in each activity data point during the training process. The learning rate is optimized for gradients.

[0137] Of course, electronic devices can also use multi-step gradient optimization to obtain the third parameter, and this application does not impose any restrictions on this.

[0138] Furthermore, the electronic device can input multiple molecular structures of the target into a meta-learning model with a third parameter to determine the molecular structure with the maximum activity value, and identify the molecular structure with the maximum activity value as the lead compound of the target.

[0139] It should be noted that in the above technical solution, the electronic device performs mixed processing of the original test set and the training set. In practice, the electronic device can also add randomly sampled bias values ​​to the original test samples in the original test set. This can be done by adding randomly sampled bias values ​​to the molecular structures in the original test samples, or by adding randomly sampled bias values ​​to the activity values ​​of the molecular structures for the target, or vice versa. Furthermore, the electronic device can also add randomly sampled bias values ​​to the training samples in the training samples. This can be done by adding randomly sampled bias values ​​to the molecular structures in the training samples, or by adding randomly sampled bias values ​​to the activity values ​​of the molecular structures for the target, or vice versa.

[0140] Optionally, if the electronic device adds a bias value obtained from random sampling to both the original test sample and the training sample, then the bias values ​​added to the two samples can be the same, or they can be different.

[0141] In summary, in this application, the electronic device can perform mixed processing on the original test set and training set data. For example, the electronic device can perform random linear mixing processing on the original test set and training set data. Here, random means that the electronic device can randomly select a corresponding training sample for any original test sample, or the electronic device can randomly select a target layer, or the electronic device can randomly select a corresponding training sample for any original test sample and randomly select a target layer. Through the technical solution provided in this application, by mixing the test set and training set, the memory of the test set by the above-mentioned meta-learning model is reduced, the generalization of the model is improved, and the model learns the relationship between molecular structure and activity value from the training set, thereby improving the performance of the meta-learning model. Consequently, the accuracy of the lead compounds determined based on this meta-learning model is higher.

[0142] Furthermore, if the electronic device performs random mixing of the data from the original test set and the training set, such as random linear mixing, this randomness can further prevent the meta-learning model from memorizing the test set, allowing the model to learn the relationship between molecular structure and activity value from the training set, thereby improving the performance of the meta-learning model. Consequently, the accuracy of the lead compounds identified based on this meta-learning model is higher.

[0143] Figure 6 A schematic diagram of a lead compound detection device 600 provided in this application embodiment is shown below. Figure 6 As shown, the device 600 includes:

[0144] The first acquisition module 610 is used to acquire a meta-learning model with initial parameters.

[0145] The second acquisition module 620 is used to acquire the activity data of the target target and multiple historical targets. For any one of the target target and multiple historical targets, the activity data of each target includes: the molecular structure of the lead compound for the target and the activity value of the molecular structure for the target.

[0146] The partitioning module 630 is used to partition the activity data of the first target to obtain the training set and the original test set corresponding to the first target. The training set includes multiple training samples, and the original test set includes multiple original test samples. The first target is any one of the multiple historical targets, or the first target is any one of the target target and the multiple historical targets.

[0147] The hybrid processing module 640 is used in the meta-learning method to select the training sample corresponding to the original test sample from the plurality of training samples for any one of the original test samples, and to perform hybrid processing on the original test sample and the training sample corresponding to the original test sample to obtain a hybrid processing result.

[0148] The optimization module 650 is used to optimize the initial parameters based on the mixing results of each of the first target points to obtain the first parameters.

[0149] The determination module 660 is used to determine a lead compound for the target target based on the first parameter, the meta-learning model, and the activity data of the target target.

[0150] Optionally, the hybrid processing module 640 is specifically used to: randomly select a training sample corresponding to the original test sample from multiple training samples.

[0151] Optionally, the hybrid processing module 640 is specifically used for: performing gradient optimization on the initial parameters of the meta-learning model based on the training set corresponding to the first target point to obtain the second parameters; for any one of the multiple original test samples, selecting the training sample corresponding to the original test sample from the multiple training samples, inputting the first molecular structure and the second molecular structure into the first neural network model respectively to obtain the first output results of the first molecular structure and the second molecular structure respectively, wherein the first molecular structure is the molecular structure of the original test sample, the second molecular structure is the molecular structure of the training sample corresponding to the original test sample, and the first neural network model is composed of the first layer to the target layer of the meta-learning model with the second parameters; performing hybrid processing on the first output results of the first molecular structure and the first output results of the second molecular structure to obtain the first hybrid processing result; inputting the first hybrid processing result into the second neural network model to obtain the second output result, wherein the second neural network model is composed of the next layer to the last layer of the target layer of the meta-learning model with the second parameters; performing hybrid processing on the activity value of the original test sample and the activity value of the training sample corresponding to the original test sample to obtain the second hybrid processing result; wherein the hybrid processing result includes: the second output result and the second hybrid processing result.

[0152] Optionally, the meta-learning model is a multi-layer deep neural network model, and the lead compound discovery device 600 further includes a selection module 670 for selecting a target layer in the multi-layer deep neural network model.

[0153] Optionally, the selection module 670 is specifically used to: randomly select a target layer in a multi-layer deep neural network model.

[0154] Optionally, the mixing module 640 is specifically used to: perform linear mixing processing on the first output result of the first molecular structure and the first output result of the second molecular structure to obtain a first mixing processing result; and perform linear mixing processing on the activity value of the original test sample and the activity value of the corresponding training sample to obtain a second mixing processing result.

[0155] Optionally, the linear mixing weight used in the linear mixing of the first output result of the first molecular structure and the first output result of the second molecular structure is the same as the linear mixing weight used in the linear mixing of the activity value of the original test sample and the activity value of the training sample corresponding to the original test sample.

[0156] Optionally, the optimization module 650 is specifically used to: determine the first loss function of the meta-learning model corresponding to the first target based on the second output result and the second hybrid processing result corresponding to the first target; and optimize the initial parameters based on the first loss function corresponding to each of the first targets to obtain the first parameters.

[0157] Optionally, the optimization module 650 is specifically used to: calculate the first gradient value of the first loss function corresponding to each of the first target points with respect to the initial parameters; optimize the initial parameters based on the first gradient value of the first loss function corresponding to each of the first target points with respect to the initial parameters to obtain the first parameters.

[0158] Optionally, the optimization module 650 is specifically used to: sum the first gradient values ​​of the first loss function corresponding to each of the first target points with respect to the initial parameters to obtain a first summation result; and optimize the initial parameters based on the first summation result to obtain the first parameters.

[0159] Optionally, the optimization module 650 is specifically used to: calculate the product of the first summation result and the preset factor to obtain the first product result; and optimize the initial parameters based on the first product result to obtain the first parameters.

[0160] Optionally, the optimization module 650 is specifically used to: calculate the difference between the initial parameters and the first product result to obtain the first parameters.

[0161] Optionally, the optimization module 650 is specifically used to: determine the first loss function of the meta-learning model corresponding to the first target based on the second output result and the second hybrid processing result corresponding to the first target; determine the second loss function corresponding to the first target based on the training set corresponding to the first target; and optimize the initial parameters based on the first loss function and the second loss function corresponding to each of the first targets to obtain the first parameters.

[0162] Optionally, the optimization module 650 is specifically used to: calculate the second gradient value of the first loss function corresponding to each of the first target points with respect to the first parameters; calculate the third gradient value of the second loss function corresponding to each of the first target points with respect to the initial parameters; and optimize the initial parameters based on the second and third gradient values ​​corresponding to each of the first target points to obtain the first parameters.

[0163] Optionally, the optimization module 650 is specifically used to: calculate the product of the third gradient value corresponding to the first target point and the gradient optimization learning rate to obtain the second product result; optimize the initial parameters according to the second gradient value and the second product result corresponding to each of the first target points to obtain the first parameters.

[0164] Optionally, the optimization module 650 is specifically used to: calculate the difference between the identity matrix and the second product result corresponding to the first target point to obtain the first difference result; optimize the initial parameters according to the second gradient value and the first difference result corresponding to each of the first target points to obtain the first parameters.

[0165] Optionally, the optimization module 650 is specifically used to: calculate the product of the second gradient value corresponding to the first target point and the first difference result to obtain the third product result; optimize the initial parameters according to the third product result corresponding to each of the first target points to obtain the first parameters.

[0166] Optionally, the optimization module 650 is specifically used to: sum the third product results corresponding to each of the first target points to obtain a second summation result; and optimize the initial parameters based on the second summation result to obtain the first parameters.

[0167] Optionally, the optimization module 650 is specifically used to: calculate the product of the second summation result and the preset factor to obtain the fourth product result; and optimize the initial parameters based on the fourth product result to obtain the first parameter.

[0168] Optionally, the optimization module 650 is specifically used to: calculate the difference between the initial parameter and the fourth product result to obtain the first parameter.

[0169] Optionally, the determination module 660 is specifically used for: performing gradient optimization on the first parameter based on the activity data of the target to obtain the third parameter; inputting multiple molecular structures of the target into a meta-learning model with the third parameter to determine the molecular structure with the maximum activity value; and determining the molecular structure with the maximum activity value as the lead compound of the target.

[0170] Optionally, the determining module 660 is specifically used to: perform single-step or multi-step gradient optimization on the first parameter based on the activity data of the target point to obtain the third parameter.

[0171] It should be understood that the device embodiments and method embodiments can correspond to each other, and similar descriptions can be referred to the method embodiments. To avoid repetition, further details will not be provided here. Specifically, Figure 6 The apparatus 600 shown can execute the above-described method embodiments, and the foregoing and other operations and / or functions of each module in the apparatus 600 are respectively for implementing the corresponding processes in the above-described methods. For the sake of brevity, they will not be described in detail here.

[0172] The apparatus 600 of this application embodiment has been described above from the perspective of functional modules in conjunction with the accompanying drawings. It should be understood that this functional module can be implemented in hardware, in software instructions, or in a combination of hardware and software modules. Specifically, the steps of the method embodiments in this application can be completed by integrated logic circuits in the processor's hardware and / or by software instructions. The steps of the method disclosed in this application embodiment can be directly embodied as being executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. Optionally, the software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps in the above method embodiments.

[0173] Figure 7 This is a schematic block diagram of the electronic device 700 provided in the embodiments of this application.

[0174] like Figure 7 As shown, the electronic device 700 may include:

[0175] The system includes a memory 710 and a processor 720. The memory 710 stores computer programs and transfers the program code to the processor 720. In other words, the processor 720 can retrieve and run the computer program from the memory 710 to implement the methods described in the embodiments of this application.

[0176] For example, the processor 720 can be used to execute the above-described method embodiments according to instructions in the computer program.

[0177] In some embodiments of this application, the processor 720 may include, but is not limited to:

[0178] General-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0179] In some embodiments of this application, the memory 710 includes, but is not limited to:

[0180] Volatile memory and / or non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0181] In some embodiments of this application, the computer program may be divided into one or more modules, which are stored in the memory 710 and executed by the processor 720 to perform the method provided in this application. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0182] like Figure 7 As shown, the electronic device may further include:

[0183] Transceiver 730, which can be connected to processor 720 or memory 710.

[0184] The processor 720 can control the transceiver 730 to communicate with other devices; specifically, it can send information or data to other devices or receive information or data sent by other devices. The transceiver 730 may include a transmitter and a receiver. The transceiver 730 may further include antennas, and the number of antennas may be one or more.

[0185] It should be understood that the various components in the electronic device are connected through a bus system, which includes a data bus, a power bus, a control bus, and a status signal bus.

[0186] This application also provides a computer storage medium storing a computer program thereon, which, when executed by a computer, enables the computer to perform the methods of the above-described method embodiments. Alternatively, embodiments of this application also provide a computer program product containing instructions that, when executed by a computer, cause the computer to perform the methods of the above-described method embodiments.

[0187] When implemented using software, it can be implemented entirely or partially as a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0188] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0189] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0190] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. For example, the functional modules in the various embodiments of this application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0191] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for discovering lead compounds, characterized in that, include: Obtain a meta-learning model with initial parameters; Acquire activity data for the target target and multiple historical targets. For any one of the target target and multiple historical targets, each activity data for the target target includes: the molecular structure of the lead compound for the target and the activity value of the molecular structure for the target. The activity data of the first target is divided to obtain the training set and the original test set corresponding to the first target. The training set includes multiple training samples, and the original test set includes multiple original test samples. The first target is any one of the multiple historical targets. In the meta-learning method, for any one of the multiple original test samples, a training sample corresponding to the original test sample is selected from the multiple training samples, and the original test sample and the training sample corresponding to the original test sample are mixed to obtain a mixed processing result. The initial parameters are optimized based on the mixed processing results of all first target points to obtain the first parameters; Based on the first parameter, the meta-learning model, and the activity data of the target, a lead compound for the target is determined.

2. The method according to claim 1, characterized in that, The step of selecting the training sample corresponding to the original test sample from the plurality of training samples includes: Randomly select a training sample corresponding to the original test sample from the plurality of training samples.

3. The method according to claim 1, characterized in that, In the meta-learning method, for any one of the plurality of original test samples, a training sample corresponding to the original test sample is selected from the plurality of training samples. The original test sample and the training sample corresponding to the original test sample are then mixed to obtain a mixed processing result, including: The initial parameters of the meta-learning model are optimized by gradient based on the training set corresponding to the first target point to obtain the second parameters; For any one of the plurality of original test samples, a training sample corresponding to the original test sample is selected from the plurality of training samples. The first molecular structure and the second molecular structure are respectively input into the first neural network model to obtain the first output result of the first molecular structure and the second molecular structure respectively. The first molecular structure is the molecular structure of the original test sample, and the second molecular structure is the molecular structure of the training sample corresponding to the original test sample. The first neural network model is composed of the first layer to the target layer of the meta-learning model with the second parameter. The first output result of the first molecular structure and the first output result of the second molecular structure are mixed to obtain the first mixed processing result; The first hybrid processing result is input into the second neural network model to obtain the second output result. The second neural network model is composed of the next layer to the last layer of the target layer of the meta-learning model with the second parameters. The activity value of the original test sample is mixed with the activity value of the training sample corresponding to the original test sample to obtain a second mixing result; The mixed processing result includes: the second output result and the second mixed processing result.

4. The method according to claim 3, characterized in that, The meta-learning model is a multi-layer deep neural network model, and the method further includes: Select the target layer in the multilayer deep neural network model.

5. The method according to claim 4, characterized in that, Selecting the target layer in the multi-layer deep neural network model includes: The target layer is randomly selected in the multi-layer deep neural network model.

6. The method according to claim 3, characterized in that, The process of mixing the first output result of the first molecular structure and the first output result of the second molecular structure to obtain the first mixed processing result includes: The first output result of the first molecular structure and the first output result of the second molecular structure are linearly mixed to obtain the first mixing result. The step of mixing the activity value of the original test sample with the activity value of the corresponding training sample to obtain a second mixing result includes: The activity value of the original test sample and the activity value of the corresponding training sample are linearly mixed to obtain the second mixing result.

7. The method according to claim 6, characterized in that, The linear mixing weights used for linearly mixing the first output result of the first molecular structure and the first output result of the second molecular structure are the same as those used for linearly mixing the activity value of the original test sample and the activity value of the training sample corresponding to the original test sample.

8. The method according to any one of claims 3-7, characterized in that, The optimization of the initial parameters based on the mixed processing results of all first target points to obtain the first parameters includes: Based on the second output result and the second hybrid processing result corresponding to the first target, determine the first loss function of the meta-learning model corresponding to the first target; The initial parameters are optimized based on the first loss function corresponding to each of the first target points to obtain the first parameters.

9. The method according to claim 8, characterized in that, The step of optimizing the initial parameters based on the first loss function corresponding to each of the first target points to obtain the first parameters includes: Calculate the first gradient value of the first loss function with respect to the initial parameters for each of the first target points; The initial parameters are optimized based on the first gradient value of the first loss function corresponding to each of the first target points with respect to the initial parameters to obtain the first parameters.

10. The method according to claim 9, characterized in that, The step of optimizing the initial parameters based on the first gradient value of the first loss function corresponding to each of the first target points with respect to the initial parameters to obtain the first parameters includes: Summing the first gradient values ​​of the first loss function corresponding to each of the first target points with respect to the initial parameters yields a first summation result. The initial parameters are optimized based on the first summation result to obtain the first parameters.

11. The method according to claim 10, characterized in that, The step of optimizing the initial parameters based on the first summation result to obtain the first parameters includes: Calculate the product of the first summation result and the preset factor to obtain the first product result; The initial parameters are optimized based on the first product result to obtain the first parameters.

12. The method according to claim 11, characterized in that, The step of optimizing the initial parameters based on the first product result to obtain the first parameters includes: The difference between the initial parameter and the first product result is calculated to obtain the first parameter.

13. The method according to any one of claims 3-7, characterized in that, The optimization of the initial parameters based on the mixed processing results of all first target points to obtain the first parameters includes: Based on the second output result and the second hybrid processing result corresponding to the first target, determine the first loss function of the meta-learning model corresponding to the first target; Based on the training set corresponding to the first target, determine the second loss function corresponding to the first target; The initial parameters are optimized based on the first loss function and the second loss function corresponding to each of the first target points to obtain the first parameters.

14. The method according to claim 13, characterized in that, The step of optimizing the initial parameters based on the first loss function and the second loss function corresponding to each of the first target points to obtain the first parameters includes: Calculate the second gradient value of the first loss function with respect to the first parameter for each of the first target points; Calculate the third gradient value of the second loss function corresponding to each of the first target points with respect to the initial parameters; The initial parameters are optimized based on the second and third gradient values ​​corresponding to each of the first target points to obtain the first parameters.

15. The method according to claim 14, characterized in that, The step of optimizing the initial parameters based on the second and third gradient values ​​corresponding to each of the first target points to obtain the first parameters includes: Calculate the product of the third gradient value corresponding to the first target point and the gradient optimization learning rate to obtain the second product result; The initial parameters are optimized based on the second gradient value and the second product result corresponding to each of the first target points to obtain the first parameters.

16. The method according to claim 15, characterized in that, The step of optimizing the initial parameters based on the second gradient values ​​and second product results corresponding to each of the first target points to obtain the first parameters includes: Calculate the difference between the identity matrix and the second product result corresponding to the first target point to obtain the first difference result; The initial parameters are optimized based on the second gradient value and the first difference result corresponding to each of the first target points to obtain the first parameters.

17. The method according to claim 16, characterized in that, The step of optimizing the initial parameters based on the second gradient values ​​and first difference results corresponding to each of the first target points to obtain the first parameters includes: Calculate the product of the second gradient value corresponding to the first target point and the first difference result to obtain the third product result; The initial parameters are optimized based on the third product results corresponding to each of the first target points to obtain the first parameters.

18. The method according to claim 17, characterized in that, The step of optimizing the initial parameters based on the third product results corresponding to each of the first target points to obtain the first parameters includes: Summing the third product results corresponding to each of the first target points yields the second summation result; The initial parameters are optimized based on the second summation result to obtain the first parameter.

19. The method according to claim 18, characterized in that, The step of optimizing the initial parameters based on the second summation result to obtain the first parameter includes: Calculate the product of the second summation result and the preset factor to obtain the fourth product result; The initial parameters are optimized based on the fourth product result to obtain the first parameter.

20. The method according to claim 19, characterized in that, The step of optimizing the initial parameters based on the fourth product result to obtain the first parameter includes: The difference between the initial parameter and the fourth product result is calculated to obtain the first parameter.

21. The method according to any one of claims 1-7, characterized in that, The step of determining a lead compound for the target target based on the first parameter, the meta-learning model, and the activity data of the target target includes: The first parameter is optimized using gradient based on the activity data of the target point to obtain the third parameter; Multiple molecular structures of the target are input into the meta-learning model with the third parameter to determine the molecular structure with the maximum activity value. The molecular structure with the highest activity value is identified as the lead compound for the target site.

22. The method according to claim 21, characterized in that, The step of performing gradient optimization on the first parameter based on the activity data of the target point to obtain the third parameter includes: The third parameter is obtained by performing single-step or multi-step gradient optimization on the first parameter based on the activity data of the target point.

23. An apparatus for discovering lead compounds, characterized in that, include: The first acquisition module is used to acquire a meta-learning model with initial parameters; The second acquisition module is used to acquire the activity data of the target target and multiple historical targets respectively. For any one of the target target and multiple historical targets, the activity data of each target includes: the molecular structure of the lead compound for the target and the activity value of the molecular structure for the target. The partitioning module is used to partition the activity data of the first target to obtain the training set and the original test set corresponding to the first target. The training set includes multiple training samples, and the original test set includes multiple original test samples. The first target is any one of the multiple historical targets. The hybrid processing module is used in the meta-learning method to select the training sample corresponding to the original test sample from the plurality of training samples for any one of the original test samples, and to perform hybrid processing on the original test sample and the training sample corresponding to the original test sample to obtain the hybrid processing result. The optimization module is used to optimize the initial parameters based on the mixing results of each of the first target points to obtain the first parameters; The determination module is used to determine a lead compound for the target target based on the first parameter, the meta-learning model, and the activity data of the target target.

24. An electronic device, characterized in that, include: A processor and a memory, the memory being used to store a computer program, the processor being used to invoke and run the computer program stored in the memory to perform the method of any one of claims 1 to 22.

25. A computer-readable storage medium, characterized in that, Used to store a computer program that causes a computer to perform the method as described in any one of claims 1 to 22.

Images