A data screening method, device, equipment, storage medium, and program product

Abstract

The embodiment of the present application discloses a data screening method, device, equipment, storage medium, and program product, the method comprising: obtaining the representation data of the target compound molecule; calling the affinity prediction model to process the representation data of the target compound molecule to obtain the affinity prediction result between the target compound molecule and the target object; wherein the affinity prediction model is obtained by self-supervised pre-training based on a first unlabeled training data set and fine-tuning based on a second labeled training data set, the first training data set includes a first sample compound molecule, the second training data set includes a second sample compound molecule, and an affinity reference label between the second sample compound molecule and the target object; the screening result of the target compound molecule is determined based on the affinity prediction result. By using the present application, compound molecule screening can be performed accurately and efficiently.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a data screening method, device, equipment, storage medium, and program product. Background Art

[0002] With the rapid development of computer technology, computer-based machine learning methods have been increasingly widely used in various fields. In the field of drug design, the task of determining whether a compound molecule is a ligand of the corresponding target can be modeled as a classification task in machine learning, that is, using a computer to determine whether the compound molecule has affinity with the corresponding target, and then screening out compound molecules with higher activity in binding to the ligand. In this way, the drug development cycle can be shortened and the cost of drug development can be controlled. However, the models trained using machine learning methods currently have low accuracy and insufficient model generalization when screening compound molecules. Summary of the invention

[0003] The embodiments of the present application provide a data screening method, device, equipment, storage medium, and program product, which can accurately and efficiently perform compound molecular screening.

[0004] On the one hand, an embodiment of the present application provides a data screening method, comprising:

[0005] Obtaining representation data of target compound molecules;

[0006] Calling an affinity prediction model to process the representation data of the target compound molecule to obtain an affinity prediction result between the target compound molecule and the target object; wherein the affinity prediction model is obtained by self-supervised pre-training based on a first unlabeled training data set and fine-tuning based on a second labeled training data set, the first training data set includes a first sample compound molecule, and the second training data set includes a second sample compound molecule and an affinity reference label between the second sample compound molecule and the target object;

[0007] The screening results of the target compound molecules are determined based on the affinity prediction results.

[0008] On the other hand, an embodiment of the present application provides a data screening device, including:

[0009] An acquisition unit, used for acquiring the representation data of the target compound molecule;

[0010] A processing unit is used to call an affinity prediction model to process the representation data of the target compound molecule to obtain an affinity prediction result between the target compound molecule and the target object; wherein the affinity prediction model is obtained by self-supervised pre-training based on a first unlabeled training data set and fine-tuning based on a second labeled training data set, the first training data set includes a first sample compound molecule, the second training data set includes a second sample compound molecule, and an affinity reference label between the second sample compound molecule and the target object; and a screening result of the target compound molecule is determined based on the affinity prediction result.

[0011] Accordingly, an embodiment of the present application provides a computer device, comprising: a memory, a processor and a network interface, wherein the processor is connected to the memory and the network interface, wherein the network interface is used to provide a network communication function, the memory is used to store program code, and the processor is used to call the program code to execute the method in the embodiment of the present application.

[0012] Accordingly, an embodiment of the present application provides a computer-readable storage medium, including: a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the method in the embodiment of the present application is implemented.

[0013] Accordingly, an embodiment of the present application provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads and executes the computer instructions from the computer-readable storage medium, so that the computer device executes the method in the embodiment of the present application.

[0014] In the embodiment of the present application, the affinity prediction model used for predicting the affinity of compound molecules is obtained after fine-tuning the prediction model obtained by self-supervised pre-training. Specifically, the initial prediction model is self-supervised pre-trained by the first unlabeled training data set to obtain a basic prediction model, and then the basic prediction model is fine-tuned by the second labeled training data set to obtain an affinity prediction model; the affinity prediction model is then called to process the representation data of the target compound molecule to obtain the affinity prediction result between the target compound molecule and the target object, and finally the screening result of the target compound molecule is determined based on the affinity prediction result. In this way, the accuracy and screening efficiency of compound molecule screening can be effectively improved, and the generalization of the model is also enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 It is a schematic diagram of the architecture of a data screening system provided in an embodiment of the present application;

[0017] Figure 2 It is a flow chart of a data screening method provided in an embodiment of the present application;

[0018] Figure 3 It is a schematic diagram of data representation of a compound molecule provided in an embodiment of the present application;

[0019] Figure 4 It is a flowchart of another data screening method provided in an embodiment of the present application;

[0020] Figure 5 This is a schematic diagram of a self-supervised label of an atom provided in an embodiment of the present application;

[0021] Figure 6 is a schematic diagram of fine-tuning a basic prediction model provided in an embodiment of the present application;

[0022] Figure 7 It is a schematic diagram of a process of self-supervised pre-training in a data screening method provided in an embodiment of the present application;

[0023] Figure 8 It is a schematic diagram of a process for obtaining an affinity prediction model in a data screening method provided in an embodiment of the present application;

[0024] Fig. 9 is a structural schematic diagram of a data screening device provided in an embodiment of the present application;

[0025] Fig.10 It is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0027] The data screening method provided in the embodiment of the present application specifically applies the machine learning technology of artificial intelligence. Machine Learning (ML) is a multi-field cross-disciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. Specializes in how computers simulate or realize human learning behavior to acquire new knowledge or skills, and reorganize the existing knowledge structure to continuously improve its performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its application covers all fields of artificial intelligence. Machine learning and deep learning generally include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning. In the data screening scheme provided in the embodiment of the present application, the affinity prediction model is called to process the representation data (such as the data in the figure) of the target compound molecule, and then the affinity prediction result between the target compound molecule and the target object is obtained, and then the screening result of the target compound molecule is determined based on the affinity prediction result. Among them, in the process of training the affinity prediction model, it is specifically divided into two stages, the first stage is self-supervised pre-training based on the first training data set without labels, and the second stage is fine-tuning based on the second training data set with labels. Compared with training the model using only labeled data sets, this solution further fine-tunes the model obtained based on self-supervised pre-training during the training process of the affinity prediction model, so that the final affinity prediction model can improve the accuracy of the affinity prediction results between the compound molecules and the target objects, thereby improving the accuracy of compound molecular screening, while also enhancing the generalization of the model.

[0028] See also Figure 1 , Figure 1 Schematic diagram of the architecture of a data screening system provided by an embodiment of the present application. Figure 1 As shown, the data screening system includes a terminal device 101, a server 102 and a database 103, wherein the terminal device 101 and the server 102 can establish a communication connection via a wired or wireless manner, and the database 103 can provide data services for the server 102.

[0029] like Figure 1As shown, the terminal device 101 can send the target compound molecule to be subjected to affinity prediction to the server 102. Before performing affinity prediction, the server 102 first obtains the representation data of the target compound molecule, and then calls the affinity prediction model to process the representation data of the target compound molecule to obtain the affinity prediction result between the target compound molecule and the target object, and finally determines the screening result of the target compound molecule based on the affinity prediction result, and can send the screening result to the terminal device 101. It is worth noting that the affinity prediction model is obtained by the server 102 obtaining a first unlabeled training data set from the database 103 to perform self-supervised pre-training on the initial prediction model to obtain a basic prediction model, and then obtaining a second labeled training data set from the database 103 to fine-tune the basic prediction model, which is conducive to improving the prediction accuracy of the affinity prediction model and enhancing the generalization of the model.

[0030] It should be noted that the terminal devices involved in the embodiments of the present application may be personal computers, laptops, smart phones, tablet computers, smart watches, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals and smart wearable devices, etc., but are not limited thereto. The server may be an independent physical server, or a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. The database may be a local database or a cloud database accessible to the server, etc., and this application does not limit this.

[0031] Based on the above description, this application embodiment provides a data screening method, see Figure 2 , Figure 2 1 is a flow chart of a data screening method provided in an embodiment of the present application. The data screening method can be executed by a computer device, which can be a terminal device or a server. For ease of understanding, the present application embodiment takes the method executed by a server as an example. Figure 2 As shown, the method may include but is not limited to the following steps.

[0032] S201: Acquire the representation data of the target compound molecule.

[0033] In the embodiments of the present application, a compound molecule is a whole composed of atoms bonded together in a certain bonding order and spatial arrangement, wherein the bonding order and spatial arrangement relationship are referred to as molecular structure, and different molecules have different molecular structures. A target compound molecule refers to a compound molecule that needs to be predicted for affinity with a target object, where the target object can refer to any target, and the representation data of the target compound molecule can specifically be a graph representation data corresponding to the target compound molecule.

[0034] Specifically, in computer science, a graph is a collection of nodes connected by a series of edges. The graph can be defined as: G(V, E), where V is the vertex set and E is the edge set. The nodes and edges of the graph can be represented by feature vectors, for example, x v It can represent the feature vector of node v, x w It can represent the feature vector of node w, e vw It can represent the eigenvector of edge vw. For compound molecules, atoms in the molecule can be represented as nodes, and chemical bonds can be represented as edges. In this way, the compound molecule can be represented by a graph data structure, such as Figure 3 As shown, Figure 3 The nicotine molecule C10H4N2 shown in (1) can be expressed as Figure 3 The graph data structure shown in (2) is a graph data structure in which atoms are nodes and chemical bonds are edges.

[0035] In one embodiment, the specific way for the server to obtain the representation data of the target compound molecule can be to obtain the properties of each atom and each chemical bond included in the target compound molecule, and then generate the representation data of the target compound molecule based on each atom, the properties of each atom, each chemical bond and the properties of each chemical bond included in the target compound molecule. Among them, the properties of atoms can be, for example, the number of charges, the number of protons, the number of neutrons, etc. of atoms, and the properties of chemical bonds can be, for example, the type of chemical bonds, the valence state of chemical bonds, etc. According to the properties of atoms and chemical bonds, the server can respectively model the characteristic vectors of nodes and the characteristic vectors of edges in the graph, so that the target compound molecule is modeled into a graph data structure, and the representation data of the target compound molecule is obtained.

[0036] In one implementation scenario, the terminal device can upload the target compound molecule to the server so that the server processes the target compound molecule and models the target compound molecule into a graph data structure, that is, obtains the representation data of the target compound molecule. Alternatively, the representation data of the target compound molecule can also be stored in a database and directly obtained by the server from the database.

[0037] S202: calling the affinity prediction model to process the representation data of the target compound molecule to obtain the affinity prediction result between the target compound molecule and the target object.

[0038] In the embodiment of the present application, after obtaining the representation data of the target compound molecule, the server can call the affinity prediction model to process the representation data of the target compound molecule to obtain the affinity prediction result between the target compound molecule and the target object. Here, the affinity prediction result refers to a label used to indicate whether the target compound molecule has affinity with the target object.

[0039] The affinity prediction model is obtained by self-supervised pre-training based on a first unlabeled training data set and fine-tuning based on a second labeled training data set, wherein the first training data set includes first sample compound molecules, and the second training data set includes second sample compound molecules and affinity reference labels between the second sample compound molecules and the target object. In one embodiment, the specific training method of the affinity prediction model can be found in Figure 4 The illustrated embodiment will not be described in detail herein.

[0040] S203: Determine the screening result of the target compound molecule based on the affinity prediction result.

[0041] In the embodiment of the present application, the server can determine the screening result of the target compound molecule based on the affinity prediction result. Specifically, if the affinity prediction result indicates that the target compound molecule has affinity with the target object, that is, it means that the target compound molecule is a ligand of the corresponding target and is an active compound molecule, then it can be determined to screen the target compound molecule; if the affinity prediction result indicates that the target compound molecule does not have affinity with the target object, that is, it means that the target compound molecule may not be a ligand of the corresponding target and is an inactive compound molecule, then the target compound molecule may not be screened.

[0042] By implementing the embodiments of the present application, the affinity prediction model can be used to predict the affinity between the target compound molecule and the target object, thereby improving the performance of virtual screening of compound molecules. The affinity prediction model can be used for virtual screening in the early stage of drug development, and can more accurately screen out active compound molecules, thereby reducing the cost of drug development. In addition, the affinity prediction model obtained by self-supervised pre-training based on the unlabeled first training data set and then fine-tuning based on the labeled second training data set also has stronger generalization.

[0043] For further information, see Figure 4 , Figure 41 is a flow chart of another data screening method provided in an embodiment of the present application. The data screening method can be executed by a computer device, which can be a terminal device or a server. For ease of understanding, the present application embodiment takes the method executed by a server as an example. Figure 4 As shown, the method may include but is not limited to the following steps.

[0044] S401: Obtaining the representation data of the target compound molecule.

[0045] In the embodiment of the present application, the execution process of step S401 can be seen in Figure 2 The specific description of step S201 is omitted here.

[0046] S402: Acquire a first training data set and a second training data set.

[0047] In an embodiment of the present application, the server can obtain a first training data set and a second training data set from a database. The first training data set is an unlabeled data set, including a first sample compound molecule, and the second training data set is a labeled data set, including a second sample compound molecule, and an affinity reference tag between the second sample compound molecule and the target object, where the affinity reference tag refers to a real affinity tag of the compound molecule measured by a chemical experiment, which is used to distinguish it from the affinity prediction tag.

[0048] S403: Perform self-supervised pre-training on the initial prediction model based on the first training data set to obtain a basic prediction model.

[0049] In an embodiment of the present application, after the server obtains the first training data set, it can perform self-supervised pre-training on the initial prediction model based on the first training data set, specifically, obtaining the pharmacophore characteristic information of each atom in the first sample compound molecule included in the first training data set, determining the self-supervised label of each atom based on the pharmacophore characteristic information of each atom, and then performing self-supervised pre-training on the initial prediction model based on the atoms and self-supervised labels included in the first sample compound molecule to obtain a basic prediction model. Among them, the pharmacophore refers to the physicochemical characteristics and spatial arrangement necessary for the ligand to be molecularly recognized by the biomacromolecule.

[0050] In one embodiment, the pharmacophore characteristic information of each atom may include any one or more of the following: whether it has a hydrogen bond acceptor characteristic, whether it has a hydrogen bond donor characteristic, whether it has a negative charge center characteristic, whether it has a positive charge center characteristic, whether it has a hydrophobic center characteristic, and whether it has an aromatic characteristic. Accordingly, each pharmacophore characteristic may correspond to a characteristic parameter. If the atom has the pharmacophore characteristic, the corresponding characteristic parameter is 1. If the atom does not have the pharmacophore characteristic, the corresponding characteristic parameter is 0. According to this rule, the self-supervisory label of each atom can be determined. For example, Figure 5 As shown, Figure 5 In the compound molecule shown, the self-supervised label of the nitrogen atom (N) is: hydrogen bond acceptor (acceptor) 0-hydrogen bond donor (donor) 1-negative charge center (NegIonizable) 0-positive charge center (PosIonizable) 0-hydrophobic center (Hydrophobe) 0-aromatic (Aromatic) 1, which means that the nitrogen atom has hydrogen bond donor characteristics and aromatic characteristics, and does not have hydrogen bond acceptor characteristics, negative charge center characteristics, positive charge center characteristics, and hydrophobic center characteristics. In this way, the self-supervised label of each atom in the first sample compound molecule can be obtained, that is, the pharmacophore characteristic label representing each atom, thereby characterizing the physicochemical characteristic information of each atom.

[0051] After determining the self-supervised label of each atom, the server can perform self-supervised pre-training on the initial prediction model based on the atoms included in the first sample compound molecule and the self-supervised label. It should be noted that the initial prediction model can be any model used for affinity prediction, such as a graph convolutional network (GCN), a random forest (Random Forest), a message passing neural network (MPNN), etc. The selection of the initial prediction model type will not affect the self-supervised pre-training process. In one embodiment, the server can perform self-supervised pre-training on the initial prediction model in a manner as follows: Figure 7 The method in the embodiment shown in the figure can be found in the following Figure 7 The corresponding embodiments are not described in detail here. Figure 7 By performing self-supervised pre-training on the initial prediction model in the manner shown, a basic prediction model can be obtained.

[0052] S404: Fine-tune the basic prediction model based on the second training data set to obtain an affinity prediction model.

[0053] In an embodiment of the present application, after the server obtains the basic prediction model, it can fine-tune the basic prediction model based on the second training data set to obtain an affinity prediction model. The fine-tuning method can be that the server samples the second training data set based on the affinity reference label of the second sample compound molecule to obtain a third training data set, and then fine-tune the basic prediction model based on the second sample compound molecule and the affinity reference label included in the third training data set to obtain the affinity prediction model.

[0054] Specifically, the affinity reference tag of the second sample compound molecule can be used to indicate whether the second sample compound molecule is an active molecule. In practical applications, for a specific target, the number of active molecules determined by chemical experiments is very small (e.g., tens to thousands), but the number of inactive molecules is usually huge, which easily leads to data imbalance, thereby affecting the accuracy and generalization of the model. Therefore, the server can sample and process the second training data set based on the affinity reference tag of the second sample compound molecule to obtain a third training data set with relatively balanced data. Based on the affinity reference tag of the second sample compound molecule, the server can determine the active molecule set and the inactive molecule set from the second training data set. The active molecule set includes active second sample compound molecules, and the inactive molecule set includes inactive second sample compound molecules. Then, based on the number of molecules included in the active molecule set, the inactive molecule set is sampled and processed to obtain the target molecule set. For example, if the number of molecules in the active molecule set is 8000, the same number (8000) of molecules can be sampled from the inactive molecule set, or based on the number of molecules in the active molecule set, the inactive molecule set is sampled according to a preset ratio (for example, the ratio of the number of active molecules / the number of inactive molecules is 0.8 to 1.6), and the sampled inactive molecules are used as the target molecule set. Finally, based on the active molecule set and the target molecule set, a third training data set for fine-tuning the basic prediction model can be determined to obtain an affinity prediction model. In one embodiment, in addition to sampling the second training data set in the above manner, a third training data set can be obtained by synthesizing a minority oversampling method. For example, assuming that the number of molecules in the active molecule set is 10 and the number of molecules in the inactive molecule set is 100, the molecules in the active molecule set can be replicated 10 times to obtain 100 active molecules, and then the 100 active molecules and 100 inactive molecules obtained by replication are determined as the third training data set, so that the purpose of balancing the data can also be achieved.

[0055] In one embodiment, by fine-tuning the basic prediction model through the third training data set, model parameters that meet the conditions can be obtained, and after the basic prediction model is updated according to the model parameters that meet the conditions, the affinity prediction model can be obtained. At this time, the performance of the affinity prediction model can be tested through the verification data set and the test data set. Optionally, the third training data set, the verification data set, and the test data set can be data sets of different molecular skeleton types, which will make the data similarity between the third training data set and the verification data set and the test data set greatly different, increase the difficulty of the experiment, and improve the reliability of the test results.

[0056] like Figure 6 As shown, a schematic diagram of fine-tuning a basic prediction model provided by an embodiment of the present application, firstly, the unbalanced data set is divided into a second training data set, a validation data set and a test data set, where the second training data set, the validation data set and the test data set can be divided according to a ratio (e.g., 8:1:1), and then the active molecule set and the inactive molecule set in the second training data set are determined, and part of the inactive molecules are randomly sampled from the inactive molecule set based on the number of active molecule sets to obtain a target molecule set equivalent to the number of active molecule sets, and finally the third training data set is determined based on the active molecule set and the target molecule set to achieve the purpose of data balance. The above process can be referred to as a process of dynamic sampling of the second training data set, and each time a complete training of the basic prediction model is performed, a dynamic sampling of the second training data set can be performed. In this way, the basic prediction model is fine-tuned, the proportion of active molecules in the training data set can be increased, and the model's learning of active molecule features can be enhanced, and the adoption process is combined with the training process, and the training data set changes dynamically with the training process, which can reduce the amount of data for each training, and there will be no information loss.

[0057] S405: calling the affinity prediction model to process the representation data of the target compound molecule to obtain the affinity prediction result between the target compound molecule and the target object.

[0058] S406: Determine the screening result of the target compound molecule based on the affinity prediction result.

[0059] In the embodiment of the present application, the execution process of step S405 and step S406 can be respectively referred to Figure 2 The specific description of step S202 and step S203 is omitted here.

[0060] By implementing the embodiments of the present application, the server can perform self-supervised pre-training on the initial prediction model based on the first unlabeled training data set to obtain a basic prediction model, and then fine-tune the basic prediction model based on the second labeled training data set to finally obtain an affinity prediction model. On the one hand, through self-supervised pre-training, the model learns more knowledge that is conducive to affinity prediction, improving the accuracy of the model. On the other hand, in the process of fine-tuning the basic prediction model, the training data set is dynamically sampled, which solves the problem of data imbalance without changing the algorithm, thereby enhancing the generalization of the model.

[0061] For further information, see Figure 7 , Figure 7 1 is a flow chart of self-supervised pre-training in a data screening method provided in an embodiment of the present application. The data screening method can be executed by a computer device, which can be a terminal device or a server. For ease of understanding, the present application embodiment takes the method executed by a server as an example. Figure 7 As shown, the method may include but is not limited to the following steps.

[0062] S701: Determine a target atom from atoms included in the first sample compound molecule based on a self-supervisory label of each atom included in the first sample compound molecule.

[0063] In an embodiment of the present application, the server can determine the target atom from the atoms included in the first sample compound molecule based on the self-supervised label of each atom included in the first sample compound molecule. The target atom refers to the atom used to predict the self-supervised label during the self-supervised pre-training process. The predicted self-supervised label here is a self-supervised pre-training task pre-set in this solution. In other possible implementation methods, other self-supervised pre-training tasks that are conducive to improving the accuracy of the model can also be set.

[0064] In one embodiment, the target atom may be determined by first determining the number of pharmacophore features of each atom based on the self-supervisory label of each atom included in the first sample compound molecule, wherein the self-supervisory label of each atom is the pharmacophore feature label of each atom, and based on the pharmacophore feature label of each atom, the number of pharmacophore features of each atom may be determined, for example, based on Figure 5 The self-supervised labels of the nitrogen atoms shown can be determined Figure 5The number of pharmacophore features of the nitrogen atom shown is 2 (hydrogen bond donor feature and aromatic feature); then the first type of atoms whose number of pharmacophore features is greater than or equal to a preset value are obtained, and the preset value may be 3, for example, then the atoms whose number of pharmacophore features is greater than or equal to 3 will be determined as the first type of atoms; then the second type of atoms are determined based on the target sampling rate from the atoms whose number of pharmacophore features is less than the preset value, and the specific method of determining the second type of atoms based on the target sampling rate may be to generate a random number s in the range of 0 to 1 for each atom whose number of pharmacophore features is less than the preset value, and then calculate the target sampling rate S using the following formula:

[0065]

[0066] Wherein, S is the target sampling rate, x is the number of pharmacophore features contained in each atom, and n is the number of pharmacophore features corresponding to the atom with the largest number of pharmacophore features in the molecule of the first sample compound. The random number s generated for each atom with a pharmacophore feature number less than a preset value is compared with the target sampling rate S calculated using the above formula: if the random number s is less than the target sampling rate S, the atom is sampled, that is, the atom is determined as a second-class atom; if the random number s is greater than or equal to the target sampling rate S, the atom is not sampled. Optionally, for each atom with a pharmacophore feature number less than a preset value, the atom with a generated random number s greater than or equal to the target sampling rate S can also be determined as a second-class atom, and the generated random number s is less than the target sampling rate S. No sampling is performed. The specific sampling rules can be set according to the actual application. Finally, the first-class atoms and the second-class atoms are both determined, and the first-class atoms and the second-class atoms are then used as target atoms.

[0067] It should be noted that in the specific implementation, since the number of atoms with aromatic characteristics is much larger than the number of atoms with other pharmacophore characteristics, in order to avoid excessive sampling of atoms with only aromatic characteristics and ignoring other pharmacophore characteristics during the sampling process, when an atom only has aromatic characteristics, the number of pharmacophore characteristics contained in the atom can be marked as 0, and the rest of the cases are counted normally, so as to achieve data balance and avoid the problem of insufficient generalization of the model during self-supervised pre-training.

[0068] S702: performing masking processing on target atoms in the first sample compound molecules to obtain masked first sample compound molecules.

[0069] In the embodiment of the present application, after determining the target atom from the first sample compound molecule, the server may mask the target atom in the first sample compound molecule, that is, mask the pharmacophore characteristic information of the target atom to obtain the masked first sample compound molecule.

[0070] S703: calling the initial prediction model to perform prediction processing on the masked first sample compound molecule to obtain a pharmacophore feature prediction label of the target atom.

[0071] In the embodiment of the present application, the initial prediction model server is called to perform prediction processing on the masked first sample compound molecule, that is, the pharmacophore feature prediction label of the target atom is obtained through the initial prediction model prediction.

[0072] S704: Adjust the model parameters of the initial prediction model based on the pharmacophore feature prediction labels of the target atoms and the self-supervised labels to obtain a basic prediction model.

[0073] In an embodiment of the present application, after obtaining the pharmacophore feature prediction label of the target atom, the server can adjust the model parameters of the initial prediction model based on the pharmacophore feature prediction label of the target atom and the self-supervised label. Specifically, the loss function of the self-supervised pre-training can be designed as the difference between the pharmacophore feature prediction label of the target atom and the self-supervised label. In the self-supervised pre-training stage, the server can minimize the loss function by adjusting the model parameters of the initial prediction model. Accordingly, when adjusting the model parameters, the loss function can be calculated to update the model parameters once after each prediction of the pharmacophore feature prediction label of a target atom. It is also possible to calculate the loss function to update the global model parameters after all the pharmacophore feature prediction labels of the target atoms in the first sample compound molecule are predicted, and finally minimize the loss function to obtain the basic prediction model.

[0074] By implementing the embodiments of the present application, the pharmacophore characteristics of atoms are combined with the self-supervised pre-training of the initial prediction model, which not only utilizes the topological similarity of the compound molecules, but also utilizes the functional similarity of the groups, so that the model learns a wider range of domain knowledge, which is conducive to improving the model effect.

[0075] For further information, see Figure 8 , Figure 8 1 is a flow chart of obtaining an affinity prediction model in a data screening method provided in an embodiment of the present application. The data screening method can be executed by a computer device, which can be a terminal device or a server. For ease of understanding, the present application embodiment is described by taking the method executed by a server as an example.

[0076] S801: The server may perform self-supervised pre-training on the initial prediction model based on the first unlabeled training data set. The self-supervised pre-training task may be a preset task that is beneficial to affinity prediction. The preset self-supervised pre-training task in this solution is an atom-based pharmacophore feature label prediction task, that is, firstly construct a self-supervised label for each atom through the pharmacophore feature information of the atoms included in the first sample compound molecule in the first training data set, and then pre-train the initial prediction model based on the atoms included in the first sample compound molecule and the self-supervised label to obtain a basic prediction model.

[0077] S802: After the server obtains the basic prediction model, it can fine-tune the basic prediction model based on the second training data set with labels to obtain the affinity prediction model. In the process of fine-tuning the model, each time the basic prediction model is trained, the second training data set can be dynamically sampled. The specific sampling process can be seen in the above Figure 4 The detailed description in the illustrated embodiment will not be repeated here.

[0078] By implementing the embodiments of the present application, on the one hand, in the process of self-supervised pre-training of the initial prediction model, the topological similarity of the compound molecules and the functional similarity of the groups are utilized, so that the basic prediction model obtained by self-supervised pre-training learns more knowledge that is conducive to affinity prediction; on the other hand, in the process of fine-tuning the basic prediction model, in view of data imbalance, a dynamic sampling method is adopted, and the sampling process is combined with the training process, so that the training data set changes with the training process, which can reduce the amount of data for each training and avoid information loss. In addition, the embodiments of the present application are flexible and versatile, and the algorithm itself will not be changed when the data set is dynamically sampled, so it is easy to expand to other models, and has a wide range of application scenarios, such as node classification tasks, link prediction tasks, etc.

[0079] For further information, see Fig. 9 , Fig. 9 is a schematic diagram of the structure of a data screening device provided in an embodiment of the present application. In one embodiment, Fig. 9 As shown, the data screening device 90 may include:

[0080] An acquisition unit 901 is used to acquire the representation data of the target compound molecule;

[0081] The processing unit 902 is used to call the affinity prediction model to process the representation data of the target compound molecule to obtain the affinity prediction result between the target compound molecule and the target object; wherein the affinity prediction model is obtained by self-supervised pre-training based on a first unlabeled training data set and fine-tuning based on a second labeled training data set, the first training data set includes a first sample compound molecule, the second training data set includes a second sample compound molecule, and an affinity reference label between the second sample compound molecule and the target object; and the screening result of the target compound molecule is determined based on the affinity prediction result.

[0082] In one embodiment, the acquisition unit 901 is also used to acquire a first training data set and a second training data set; the processing unit 902 is also used to perform self-supervised pre-training on the initial prediction model based on the first training data set to obtain a basic prediction model; and fine-tune the basic prediction model based on the second training data set to obtain an affinity prediction model.

[0083] In one embodiment, the acquisition unit 901 is further used to obtain the pharmacophore characteristic information of each atom in the first sample compound molecule included in the first training data set; the processing unit 902 is further used to determine the self-supervised label of each atom based on the pharmacophore characteristic information of each atom; and the initial prediction model is self-supervised pre-trained based on the atoms included in the first sample compound molecule and the self-supervised labels to obtain a basic prediction model.

[0084] In one embodiment, the processing unit 902 is further used to perform masking processing on the target atom in the first sample compound molecule to obtain the masked first sample compound molecule; call the initial prediction model to perform prediction processing on the masked first sample compound molecule to obtain the pharmacophore feature prediction label of the target atom; adjust the model parameters of the initial prediction model based on the pharmacophore feature prediction label of the target atom and the self-supervised label to obtain the basic prediction model.

[0085] In one embodiment, the processing unit 902 is also used to determine the number of pharmacophore features of each atom based on the self-supervised label of each atom included in the first sample compound molecule; obtain the first type of atoms whose pharmacophore feature number is greater than or equal to a preset value; determine the second type of atoms from the atoms whose pharmacophore feature number is less than the preset value based on the target sampling rate; and use the first type of atoms and the second type of atoms as target atoms.

[0086] In one embodiment, the processing unit 902 is further used to perform sampling processing on the second training data set based on the affinity reference label of the second sample compound molecule to obtain a third training data set; and fine-tune the basic prediction model based on the second sample compound molecules and the affinity reference label included in the third training data set to obtain an affinity prediction model.

[0087] In one embodiment, the processing unit 902 is further used to determine an active molecule set and an inactive molecule set from the second training data set based on the affinity reference tags of the second sample compound molecules, the active molecule set including active second sample compound molecules, and the inactive molecule set including inactive second sample compound molecules; perform sampling processing on the inactive molecule set based on the number of molecules included in the active molecule set to obtain a target molecule set; and determine a third training data set based on the active molecule set and the target molecule set.

[0088] In one embodiment, the acquisition unit 901 is also used to acquire the properties of each atom and each chemical bond included in the target compound molecule; the processing unit 902 is also used to generate representation data of the target compound molecule based on each atom, the properties of each atom, each chemical bond and the properties of each chemical bond included in the target compound molecule.

[0089] In one embodiment, the pharmacophore characteristic information includes any one or more of the following: whether it has a hydrogen bond acceptor characteristic, whether it has a hydrogen bond donor characteristic, whether it has a negative charge center characteristic, whether it has a positive charge center characteristic, whether it has a hydrophobic center characteristic, and whether it has an aromatic characteristic.

[0090] It should be noted that Fig. 9 The contents not mentioned in the corresponding embodiments and the specific implementation methods of each step can be found in Figure 2-Figure 8 The illustrated embodiments and the aforementioned contents will not be described in detail here.

[0091] For further information, see Fig.10 , Fig.10It is a structural diagram of a computer device 100 provided in an embodiment of the present application. The computer device may include: a network interface 1001, a memory 1002 and a processor 1003, the network interface 1001, the memory 1002 and the processor 1003 are connected through one or more communication buses, and the communication bus is used to realize the connection and communication between these components. The network interface 1001 may include a standard wired interface and a wireless interface (such as a WIFI interface). The memory 1002 may include a volatile memory (volatile memory), such as a random access memory (random-access memory, RAM); the memory 1002 may also include a non-volatile memory (non-volatile memory), such as a flash memory (flash memory), a solid-state drive (solid-state drive, SSD), etc.; the memory 1002 may also include a combination of the above-mentioned types of memory. The processor 1003 may be a central processing unit (CPU). The processor 1003 may further include a hardware chip. The above-mentioned hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), etc. The PLD may be a field-programmable gate array (FPGA), a generic array logic (GAL), or the like.

[0092] Optionally, the memory 1002 is also used to store program instructions, and the processor 1003 can also call the program instructions to implement relevant methods and steps in this application.

[0093] In one embodiment, the processor 1003 calls the program instructions stored in the memory 1002 to achieve: obtaining the representation data of the target compound molecule; calling the affinity prediction model to process the representation data of the target compound molecule to obtain the affinity prediction result between the target compound molecule and the target object; wherein the affinity prediction model is obtained by self-supervised pre-training based on a first unlabeled training data set and fine-tuning based on a second labeled training data set, the first training data set includes a first sample compound molecule, the second training data set includes a second sample compound molecule, and an affinity reference label between the second sample compound molecule and the target object; and determining the screening result of the target compound molecule based on the affinity prediction result.

[0094] In one embodiment, the processor 1003 may also call the program instructions to implement: obtaining a first training data set and a second training data set; performing self-supervised pre-training on the initial prediction model based on the first training data set to obtain a basic prediction model; and fine-tuning the basic prediction model based on the second training data set to obtain an affinity prediction model.

[0095] In one embodiment, the processor 1003 may also call the program instructions to achieve: obtaining the pharmacophore characteristic information of each atom in the first sample compound molecule included in the first training data set; determining the self-supervised label of each atom based on the pharmacophore characteristic information of each atom; and performing self-supervised pre-training on the initial prediction model based on the atoms included in the first sample compound molecule and the self-supervised labels to obtain a basic prediction model.

[0096] In one embodiment, the processor 1003 may also call the program instructions to implement: masking the target atom in the first sample compound molecule to obtain the masked first sample compound molecule; calling the initial prediction model to perform prediction processing on the masked first sample compound molecule to obtain the pharmacophore characteristic prediction label of the target atom; adjusting the model parameters of the initial prediction model based on the pharmacophore characteristic prediction label of the target atom and the self-supervised label to obtain the basic prediction model.

[0097] In one embodiment, the processor 1003 may also call the program instructions to achieve: determining the number of pharmacophore features of each atom based on the self-supervisory label of each atom included in the first sample compound molecule; obtaining the first type of atoms whose pharmacophore feature number is greater than or equal to a preset value; determining the second type of atoms from the atoms whose pharmacophore feature number is less than a preset value based on the target sampling rate; and using the first type of atoms and the second type of atoms as target atoms.

[0098] In one embodiment, the processor 1003 can also call the program instructions to achieve: sampling and processing the second training data set based on the affinity reference label of the second sample compound molecule to obtain a third training data set; fine-tuning the basic prediction model based on the second sample compound molecules and the affinity reference label included in the third training data set to obtain an affinity prediction model.

[0099] In one embodiment, the processor 1003 may also call the program instructions to achieve: determining an active molecule set and an inactive molecule set from a second training data set based on the affinity reference tags of the second sample compound molecules, the active molecule set including active second sample compound molecules, and the inactive molecule set including inactive second sample compound molecules; sampling the inactive molecule set based on the number of molecules included in the active molecule set to obtain a target molecule set; and determining a third training data set based on the active molecule set and the target molecule set.

[0100] In one embodiment, the processor 1003 may also call the program instructions to achieve: obtaining the properties of each atom and the properties of each chemical bond included in the target compound molecule; generating representation data of the target compound molecule based on each atom, the properties of each atom, each chemical bond and the properties of each chemical bond included in the target compound molecule.

[0101] In one embodiment, the pharmacophore characteristic information includes any one or more of the following: whether it has a hydrogen bond acceptor characteristic, whether it has a hydrogen bond donor characteristic, whether it has a negative charge center characteristic, whether it has a positive charge center characteristic, whether it has a hydrophobic center characteristic, and whether it has an aromatic characteristic.

[0102] It should be understood that the principles and beneficial effects of the computer device 100 described in the embodiments of the present application for solving the problem are similar to those of the present application. Figure 2-Figure 8 The principles and beneficial effects of the illustrated embodiment and the aforementioned contents in solving the problems are similar, and for the sake of brevity, they will not be repeated here.

[0103] In addition, the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method provided in the aforementioned embodiment is implemented.

[0104] The embodiment of the present application also provides a computer program product or a computer program, which includes a computer instruction stored in a computer-readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method provided in the above embodiment.

[0105] The steps in the method of the embodiment of the present application can be adjusted in order, combined and deleted according to actual needs.

[0106] The units in the device of the embodiment of the present application can be merged, divided and deleted according to actual needs.

[0107] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The above-mentioned storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0108] The above disclosure is only part of the embodiments of the present application, which certainly cannot be used to limit the scope of rights of the present application. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and equivalent changes made according to the claims of the present application are still within the scope covered by the present application.

Claims

1. A data screening method, characterized in that: The method comprises: Obtaining a first training data set and a second training data set; Performing self-supervised pre-training on the initial prediction model based on the first training data set to obtain a basic prediction model; Fine-tuning the basic prediction model based on the second training data set to obtain an affinity prediction model; Obtaining representation data of target compound molecules; Calling the affinity prediction model to process the representation data of the target compound molecule to obtain an affinity prediction result between the target compound molecule and the target object; Determining the screening result of the target compound molecule based on the affinity prediction result; The self-supervised pre-training of the initial prediction model based on the first training data set to obtain a basic prediction model includes: Acquire pharmacophore characteristic information of each atom in the first sample compound molecule included in the first training data set; determine a self-supervised label of each atom based on the pharmacophore characteristic information of each atom; Determining the number of pharmacophore features of each atom based on the self-supervised label of each atom included in the first sample compound molecule; Acquire first-class atoms whose pharmacophore feature number is greater than or equal to a preset value; determine second-class atoms from atoms whose pharmacophore feature number is less than the preset value based on a target sampling rate; and use the first-class atoms and the second-class atoms as target atoms; Performing a masking process on the target atom in the first sample compound molecule to obtain a masked first sample compound molecule; calling an initial prediction model to perform a prediction process on the masked first sample compound molecule to obtain a pharmacophore feature prediction label of the target atom; The model parameters of the initial prediction model are adjusted based on the pharmacophore feature prediction labels of the target atoms and the self-supervised labels to obtain a basic prediction model.

2. The method according to claim 1, characterized in that The fine-tuning of the basic prediction model based on the second training data set to obtain the affinity prediction model includes: Performing sampling processing on the second training data set based on the affinity reference tag of the second sample compound molecule to obtain a third training data set; The basic prediction model is fine-tuned based on the second sample compound molecules and affinity reference labels included in the third training data set to obtain an affinity prediction model.

3. The method according to claim 2, characterized in that The sampling process of the second training data set based on the affinity reference tag of the second sample compound molecule to obtain a third training data set includes: Determining an active molecule set and an inactive molecule set from the second training data set based on the affinity reference tags of the second sample compound molecules, the active molecule set including active second sample compound molecules, and the inactive molecule set including inactive second sample compound molecules; Based on the number of molecules included in the active molecule set, the inactive molecule set is sampled to obtain a target molecule set; A third training data set is determined based on the active molecule set and the target molecule set.

4. The method according to claim 1, characterized in that The step of obtaining the representation data of the target compound molecule comprises: Obtain the properties of each atom and each chemical bond included in the target compound molecule; Based on each atom included in the target compound molecule, the attribute of each atom, each chemical bond and the attribute of each chemical bond, the representation data of the target compound molecule is generated.

5. The method according to claim 1, characterized in that The pharmacophore characteristic information includes any one or more of the following: whether it has a hydrogen bond acceptor characteristic, whether it has a hydrogen bond donor characteristic, whether it has a negative charge center characteristic, whether it has a positive charge center characteristic, whether it has a hydrophobic center characteristic, and whether it has an aromatic characteristic.

6. A data screening device, characterized in that: The device comprises: An acquisition unit, used to acquire a first training data set and a second training data set; A processing unit, configured to perform self-supervised pre-training on the initial prediction model based on the first training data set to obtain a basic prediction model; and fine-tune the basic prediction model based on the second training data set to obtain an affinity prediction model; The acquisition unit is also used to acquire the representation data of the target compound molecule; The processing unit is further used to call the affinity prediction model to process the representation data of the target compound molecule to obtain the affinity prediction result between the target compound molecule and the target object; and determine the screening result of the target compound molecule based on the affinity prediction result; The acquisition unit is further used to acquire pharmacophore characteristic information of each atom in the first sample compound molecule included in the first training data set; The processing unit is also used to determine the self-supervised label of each atom based on the pharmacophore feature information of each atom; determine the number of pharmacophore features of each atom based on the self-supervised label of each atom included in the first sample compound molecule; obtain the first type of atoms whose pharmacophore feature number is greater than or equal to a preset value; determine the second type of atoms from the atoms whose pharmacophore feature number is less than the preset value based on the target sampling rate; use the first type of atoms and the second type of atoms as target atoms; mask the target atoms in the first sample compound molecule to obtain the masked first sample compound molecule; call the initial prediction model to perform prediction processing on the masked first sample compound molecule to obtain the pharmacophore feature prediction label of the target atom; adjust the model parameters of the initial prediction model based on the pharmacophore feature prediction label of the target atom and the self-supervised label to obtain the basic prediction model.

7. A computer device, characterized in that: It includes a memory, a processor and a network interface, the processor is connected to the memory and the network interface, wherein the network interface is used to provide a network communication function, the memory is used to store program code, and the processor is used to call the program code to execute the method described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

9. A computer program product, characterized in that The computer program product comprises a computer program or computer instructions, and when the computer program or computer instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images