Patent Application
20230274797
The present disclosure relates to the technical field of biological medicine, in particular to a method for training molecular binding models, a method for screening molecules, apparatuses, a computer device and a storage medium.
As technology advances, more and more devices can use activity prediction models to predict the binding activity of compounds after virtual binding of molecules, so that some molecules can be screened out for actual drug binding experiments based on the predicted binding activity of the compounds.
When predicting the binding activity of the compounds after virtual binding of the molecules, the activity prediction model is usually based on the chemical property characteristics of the molecules extracted by a feature extraction model or the structural characteristics of the molecule itself to perform prediction.
An embodiment of the present disclosure provides a method for training molecular binding models executed by a computer device. The method includes obtaining protein feature information of sample protein molecules and molecular feature information of sample alternative molecules; inputting the protein feature information and the molecular feature information into a molecular binding model, and using the molecular binding model to determine binding activity feature information, embedding feature information and eutectic feature information between the sample protein molecule and the sample alternative molecule, the binding activity feature information characterizing activities of the sample protein molecule and the sample alternative molecule after virtual binding, the embedding feature information characterizing a degree of binding between the sample protein molecule and the sample alternative molecule, and the eutectic feature information characterizing whether a eutectic structure exists between the sample protein molecule and the sample alternative molecule; determining a training loss of the molecular binding model based on the binding activity feature information, the embedding feature information and the eutectic feature information; and outputting the molecular binding model as a trained molecular binding model when the training loss of the molecular binding model meets a training target.
Another embodiment of the present disclosure provides a computing device. The computing device includes one or more processors and one or more memories, the one or more memories storing at least one computer program, the at least one computer program being loaded and executed by the one or more processors to implement a method for training molecular binding models. The method includes obtaining protein feature information of sample protein molecules and molecular feature information of sample alternative molecules; inputting the protein feature information and the molecular feature information into a molecular binding model, and using the molecular binding model to determine binding activity feature information, embedding feature information and eutectic feature information between the sample protein molecule and the sample alternative molecule, the binding activity feature information characterizing activities of the sample protein molecule and the sample alternative molecule after virtual binding, the embedding feature information characterizing a degree of binding between the sample protein molecule and the sample alternative molecule, and the eutectic feature information characterizing whether a eutectic structure exists between the sample protein molecule and the sample alternative molecule; determining a training loss of the molecular binding model based on the binding activity feature information, the embedding feature information and the eutectic feature information; and outputting the molecular binding model as a trained molecular binding model when the training loss of the molecular binding model meets a training target.
Another embodiment of the present disclosure provides a non-transitory computer readable storage medium, storing at least one computer program, the at least one computer program being loaded and executed by a processor to implement a method for training molecular binding models. The method includes obtaining protein feature information of sample protein molecules and molecular feature information of sample alternative molecules; inputting the protein feature information and the molecular feature information into a molecular binding model, and using the molecular binding model to determine binding activity feature information, embedding feature information and eutectic feature information between the sample protein molecule and the sample alternative molecule, the binding activity feature information characterizing activities of the sample protein molecule and the sample alternative molecule after virtual binding, the embedding feature information characterizing a degree of binding between the sample protein molecule and the sample alternative molecule, and the eutectic feature information characterizing whether a eutectic structure exists between the sample protein molecule and the sample alternative molecule; determining a training loss of the molecular binding model based on the binding activity feature information, the embedding feature information and the eutectic feature information; and outputting the molecular binding model as a trained molecular binding model when the training loss of the molecular binding model meets a training target.
To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the following clearly and completely describes the technical solutions in the embodiments of the present disclosure with reference to the accompanying drawings in the embodiments of the present disclosure.
Some of the terms used in the embodiments of the present disclosure are explained below to facilitate understanding by those skilled in the art.
(1) 50% Inhibitory Concentrations IC50 and pIC50:
IC50 may also be referred to as the 50% inhibition rate, and the concentration of drug or inhibitor required to inhibit half of a specified biological process. There is a mathematical relationship between pIC50 and IC50, with pIC50 being the opposite number of the logarithm of IC50.
Protein molecules are the material basis of life, organic macromolecules, the basic organic substances that make up cells and the main bearers of vital activities, and amino acid molecules are the basic constituent units of the protein molecules.
Molecules are composed of atoms which are bound together in a certain bonding order and spatial arrangement. The bonding order and spatial arrangement relationship are the molecular structures, which can also be referred to as small molecules in the embodiments of the present disclosure.
In the process of drug research and development, drug screening is one of the main ways to discover lead compounds. Physical drug screening requires the construction of a large-size compound library, the extraction or cultivation of a large number of experimentally necessary target enzymes or target cells, and the support of complex devices, so physical drug screening requires a huge capital investment. Virtual drug screening allows for preliminary screening of drug compounds prior to physical screening, thus significantly saving the cost.
In the embodiments of the present disclosure, alternative molecules refer to candidate molecules for virtual drug screening.
The embodiments of the present disclosure relate to the Artificial Intelligence (AI) technology. The design is based on the Computer Vision (CV) technology and the Machine Learning (ML) technology in the AI technology.
The following is a brief description of the fields of application of methods for training molecular binding models according to an embodiment of the present disclosure.
As technology advances, more and more devices can use activity prediction models to predict the binding activity of compounds after virtual binding of molecules, so that some molecules can be screened for realistic drug binding experiments based on the predicted binding activity of the compounds, avoiding the situation where some unnecessary drug binding experiments are performed, resulting in a waste of resources.
When predicting the binding activity of the compounds after virtual binding of the molecules, the activity prediction model is usually based on the chemical property characteristics of the molecules extracted by a feature extraction model or the structural characteristics of the molecule itself to perform prediction on the binding activity between two molecules.
However, in the molecular binding process, the two molecules are not simply put together, and there may be unpredictable interactions between the molecules. Only from the perspective of the molecules themselves, the prediction of the binding activity of a compound after virtual binding of two molecules is less accurate, resulting in low accuracy for screening molecules. As a result, some molecules that are necessary for binding are not screened out for drug binding experiments, while some molecules that are not necessary for binding are screened out for drug binding experiments, which will not only cause waste of resources, but also affect the progress of drug research and development.
In order to solve the problem that the accuracy for screening molecules is low, an embodiment of the present disclosure provides a method for training molecular binding models. Please refer to
For iterative training with many times, the following operations are performed respectively: use the to-be-trained molecular binding model to determine, based on the protein feature information and the molecular feature information, the binding activity feature information, embedding feature information and eutectic feature information between the sample protein molecule and the sample alternative molecule, the binding activity feature information characterizing the activity of the sample protein molecule and the sample alternative molecule after virtual binding, the embedding feature information characterizing the degree of binding between the sample protein molecule and the sample alternative molecule, the eutectic feature information characterizing whether a eutectic structure exists between the sample protein molecule and the sample alternative molecule, and the to-be-trained molecular binding model being configured to, based on the binding activity feature information, the embedding feature information and the eutectic feature information, determine the training loss of the to-be-trained molecular binding model.
In the embodiment of the present disclosure, after obtaining the respective protein feature information of each sample protein molecule and the respective molecular feature information of each sample alternative molecule, the binding activity feature information, the embedding feature information and the eutectic feature information between the sample protein molecule and the sample alternative molecule are further obtained based on each protein feature information and each molecular feature information; based on the binding activity feature information, the embedding feature information and the eutectic feature information, the molecular binding model is trained; the molecular binding model is trained from the perspective of multiple feature information, so that when predicting the binding activity of molecules after virtual binding, an activity prediction model takes into account not only the chemical property characteristics of molecules contained in the protein feature information and the molecular feature information, or the structural characteristics of the molecules themselves, but also the intermolecular interactions such as intermolecular embedding characteristics and eutectic characteristics. This makes the activity value predicted by the activity prediction model based on the molecular binding model more accurate.
The following illustrates the application scenario of the method for training molecular binding models according to the present disclosure.
Please refer to
The client 101 generally refers to a device that can instruct the server 102 to train molecular binding models or perform molecular screening, for example, a terminal device, a third-party application accessible by the terminal device or a web page accessible by the terminal device, etc. The terminal device includes, but is not limited to, a mobile phone, a computer, a smart medical device, a smart electric appliance, etc. The server 102 refers generally to a device that can train molecular binding models or perform molecular screening, for example, a terminal device or a server, etc. The server includes, but is not limited to, a cloud server, a local server or an associated third-party server, etc. Both the client 101 and the server 102 may use cloud computing so as to reduce the occupation of local computing resources; cloud storage may likewise be used to reduce the occupation of local storage resources.
As an embodiment, the client 101 and the server 102 may be the same device, and in the embodiment of the present disclosure, the client 101 and the server 102 are introduced as different devices as an example.
Based on
Please refer to
S301: Obtain protein feature information of sample protein molecules and molecular feature information of sample alternative molecules.
The server may receive the protein feature information of the sample protein molecule and the molecular feature information of the sample alternative molecule sent by other devices, and may also perform feature extraction processing on the sample protein molecule and the sample alternative molecule separately to obtain the protein feature information of the sample protein molecule and the molecular feature information of the sample alternative molecule, etc., which is not specifically limited. For example,
In some embodiments, the topological structure of the sample protein molecule and the sample alternative molecule can be effectively introduced by transforming the structural information of the sample protein molecule and the sample alternative molecule into an adjacency matrix as an input of a molecular binding model. The topological structure of a compound can influence the features of the whole compound, so the molecular binding model obtains the feature information of the sample protein molecule and the sample alternative molecule by means of the topological structure of the sample protein molecule and the sample alternative molecule, such as a feature matrix representation.
The embodiment of the present disclosure is introduced by taking that a server uses a feature extraction model to extract the protein feature information of the sample protein molecule and the molecular feature information of the sample alternative molecule as an example.
In some embodiments, the obtain protein feature information of sample protein molecules and molecular feature information of sample alternative molecules may include:
In some embodiments, the server may first obtain a training sample set, the training sample set including a plurality of training samples, each training sample including sample protein molecules and sample alternative molecules.
The server may extract, based on a database, a plurality of sample protein molecules and a plurality of sample alternative molecules which have undergone drug binding experiments to obtain the training sample set; the plurality of sample protein molecules and the plurality of sample alternative molecules which have undergone drug binding experiments may also be found in web resources, thereby obtaining the training sample set, with no specific limitations on a method for obtaining training sample set.
After obtaining the training sample set, the server can use the feature extraction model to perform feature extraction processing on each training sample to obtain protein feature information of the sample protein molecule and molecular feature information of the sample alternative molecule in each training sample. The feature extraction model is configured to perform feature extraction processing on the molecule to obtain the structural information of the molecule. A feature extraction model server may input respective three-dimensional structural diagrams of the sample protein molecule and the sample alternative molecule into the feature extraction model to perform feature extraction processing on each training sample respectively; may also input respective multi-perspective two-dimensional planar structural diagrams of the sample protein molecule and the sample alternative molecule into the feature extraction model to perform feature extraction processing, etc. on each training sample, without specific limitation.
As an embodiment, when the server uses the feature extraction model to perform feature extraction processing on each training sample respectively, an adjacency matrix may be used in performing calculation to introduce the process of performing feature extraction processing on one training sample, and the process of performing feature extraction processing on other training samples is similar and will not be repeated here.
The server determines an adjacency matrix of the sample protein molecule based on the sample protein molecule in the training sample; please refer to
The server determines an adjacency matrix of the sample alternative molecule based on the sample alternative molecule in the training sample; please refer to
The server performs feature extraction processing on the adjacency matrix of the sample protein molecule and the adjacency matrix of the sample alternative molecule, respectively; please refer to
S302: Input the protein feature information and the molecular feature information into a molecular binding model, and use the molecular binding model to determine the binding activity feature information, the embedding feature information and the eutectic feature information between the sample protein molecule and the sample alternative molecule; determine a training loss of the molecular binding model based on the binding activity feature information, the embedding feature information and the eutectic feature information; and output the molecular binding model as a trained molecular binding model when the training loss of the molecular binding model meets a training target; the trained molecular binding model being configured to determine binding activity feature information between a target protein molecule and a target alternative molecule to predict the binding activity of a compound after virtual binding of the target protein molecule and the target alternative molecule.
After obtaining the protein feature information of the sample protein molecule and the molecular feature information of the sample alternative molecule, the server may perform iterative training for many times on the to-be-trained molecular binding model based on the protein feature information and the molecular feature information, the process of each iterative training is similar, the process of one iterative training is introduced below, and please refer to
S501: Use the to-be-trained molecular binding model to determine, based on the protein feature information and the molecular feature information, the binding activity feature information, the embedding feature information and the eutectic feature information between the sample protein molecule and the sample alternative molecule.
After obtaining the protein feature information and the molecular feature information, the server can use the to-be-trained molecular binding model to predict the binding activity feature information between the sample protein molecule and the sample alternative molecule, the binding activity feature information being configured to predict the binding activity feature of the sample protein molecule and the sample alternative molecule after virtual binding. At the same time, the server can also use the to-be-trained molecular binding model to predict the embedding feature information and the eutectic feature information between the sample protein molecule and the sample alternative molecule, and train the molecular binding model based on the embedding feature information and the eutectic feature information, so that the molecular binding model can have the performance of identifying the embedding relationship and the eutectic structure between molecules after virtual binding, so that when the binding activity feature information between molecules after virtual binding is predicted, more accurate binding activity feature information can be predicted, and when the activity value is predicted by the activity prediction model based on the binding activity feature information obtained by the molecular binding model, more accurate binding activity can be obtained .
The process of predicting the embedding feature information and the eutectic feature information between the sample protein molecule and the sample alternative molecule using the to-be-trained molecular binding model is described below.
The server can predict, based on the protein feature information and the molecular feature information, the binding distance between the various sample amino acid molecules contained in the sample protein molecule, and the various sample alternative atoms contained in the sample alternative molecule, after virtual binding of the sample protein molecule and the sample alternative molecule. In some embodiments, the binding distance may be a distance between each sample amino acid molecule and each sample alternative atom calculated from the protein feature information and the molecular feature information. The server may determine the embedding feature information and the eutectic feature information between the sample protein molecule and the sample alternative molecule based on the various binding distances obtained.
As an embodiment, a residue of a specified sample amino acid molecule can bind to a cavity of an sample alternative atom to form an embedding relationship; therefore, the server can determine the embedding feature information based on various binding distances, and then introduce supervision information for the to-be-trained molecular binding model by means of graph matching, so that the trained molecular binding model has the performance of identifying the embedding feature between molecules, and the accuracy for predicting, based on the trained molecular binding model, the binding activity between molecules after virtual binding is improved.
After obtaining the binding distance between each sample amino acid molecule contained in the sample protein molecule and each sample alternative atom contained in the sample alternative molecule, the server can determine, based on the binding distance with a minimum value between the specified sample amino acid molecule and each sample alternative atom, the embedding feature information between the sample protein molecule and the sample alternative molecule. The specified sample amino acid molecule may be a sample amino acid molecule having a targeting property in the sample protein molecule, or a sample amino acid molecule specified according to a scenario, etc., without specific limitation. Training is stabilized to a minimum value by means of the binding distance with a minimum value between the specified sample amino acid molecule and each sample alternative atom such that the residue of the sample amino acid molecule binds to the cavity of the corresponding sample alternative atom to form an embedding relationship. Please refer to
As an embodiment, when the sample amino acid molecule matches the sample alternative atom, then there is a eutectic structure between the sample amino acid molecule and the sample alternative atom, i.e., the eutectic structure is reflected in the matching of the sample alternative atom and the sample amino acid, therefore the server can determine the eutectic feature information based on the various binding distances and then introduce supervision information for the to-be-trained molecular binding model by means of neighborhood consensus, so that the trained molecular binding model has the performance of identifying intermolecular eutectic features and the accuracy of predicting, based on the trained molecular binding model, the binding activity between molecules after virtual binding is improved.
After obtaining the binding distance between each sample amino acid molecule contained in the sample protein molecule and each sample alternative atom contained in the sample alternative molecule, the server can determine, based on each sample amino acid molecule, each sample alternative atom, and a binding distance between each sample amino acid molecule and the sample alternative atom, the eutectic feature information between the sample protein molecule and the sample alternative molecule, where the eutectic feature information can be represented in the form of an adjacency matrix. Training is stabilized to be close to a eutectic target by means of the binding distance between the sample alternative atom and the sample amino acid molecule which are matched such that between the sample alternative atom and the sample amino acid molecule which are matched, a eutectic structure is formed; please refer to
S502: Determine the training loss of the to-be-trained molecular binding model based on the binding activity feature information, the embedding feature information and the eutectic feature information.
After obtaining the binding activity feature information, the embedding feature information and the eutectic feature information, the server may determine the training loss of the to-be-trained molecular binding model based on the binding activity feature information, the embedding feature information and the eutectic feature information, and train the to-be-trained molecular binding model based on the training loss.
The server may perform weighted fusion processing on the binding activity feature information, the embedding feature information and the eutectic feature information to obtain fused feature information, and determine the training loss of the to-be-trained molecular binding model by means of an error value between the fused feature information and the training target.
The server may also determine a first training loss of the to-be-trained molecular binding model based on a first error value between the embedding feature information and an embedding target, determine a second training loss of the to-be-trained molecular binding model based on a second error value between the eutectic feature information and a eutectic target, determine a third training loss of the to-be-trained molecular binding model based on a third error value between the binding activity feature information and an activity target, and determine the training loss of the to-be-trained molecular binding model based on the first training loss, the second training loss and the third training loss which are obtained. The server may determine the training loss of the to-be-trained molecular binding model based on the weighted sum of the first training loss, the second training loss and the third training loss, and may also directly determine the first training loss, the second training loss and the third training loss as the training loss of the to-be-trained molecular binding model, etc.
As an embodiment, the server may determine, by means of a graph matching model, a first error value between the binding distance, characterized by the embedding feature information, between the specified sample amino acid molecule and the corresponding sample alternative atom, and a reference binding distance. The server determines, based on the first error value, the first training loss of the to-be-trained molecular binding model. The reference binding distance may be obtained from a PDBbind dataset.
As an embodiment, it is easier to train a neighborhood consensus model as it has a more single function and more training data, and the accuracy of a matching matrix obtained using the trained neighborhood consensus model is already at a high level. At this moment, the trained neighborhood consensus model can be used to determine a eutectic target. The server can use the trained neighborhood consensus model to match each sample amino acid molecule contained in the sample protein molecule with each sample alternative atom contained in the sample alternative molecule to obtain matching feature information between the sample protein molecule and the sample alternative molecule as the eutectic target, the matching feature information characterizing the matching distance between each sample amino acid molecule and each sample alternative atom. The server may use a cross entropy function to determine a second error value between the eutectic feature information and the eutectic target, and uses the second error value as the second training loss of the to-be-trained molecular binding model. The trained neighborhood consensus model may be trained and obtained based on data in the PDBbind dataset, and the cross entropy function may be replaced with other functions, without specific limitation.
As an embodiment, in the various training samples contained in the training sample set, in addition to the sample protein molecule and the sample alternative molecule, a reference activity value of the sample protein molecule and the sample alternative molecule after virtual binding is further included, the reference activity value characterizing the true activity of the sample protein molecule and the sample alternative molecule after the drug binding experiment. If the sample protein molecule and the sample alternative molecule are obtained based on the drug binding experiment, the reference activity value is obtained based on the experimental results of the drug binding experiment of the sample protein molecule and the sample alternative molecule. If the sample protein molecule and the sample alternative molecule are obtained from a web resource, then the reference activity value is also obtained from the corresponding web resource, etc., without specific limitation. The training sample set can be obtained from the PDBbind dataset.
The server can use the activity prediction model to predict, based on the binding activity feature information, the sample activity value after virtual binding of the sample protein molecule and the sample alternative molecule, and determine, based on the error value between the sample activity value and a corresponding reference activity value, the third training loss of the to-be-trained molecular binding model.
In some embodiments, prediction of the sample activity value includes: for a pair of given sample protein molecule and sample alternative molecule, the process of abstracting the sample protein molecule into a graph takes amino acids as vertices, uses the pdb structure file of the protein, takes the protein side chain heavy atom CB (CA is selected if CB is not present) therein as the center of the amino acid, calculates the Euclidean distance between the amino acid centers, and then by some normalized processing, obtains an adjacency matrix representing the protein structure. For the sample alternative molecule, the adjacency matrix thereof is also obtained in the same way by using atoms directly as vertices. Two graphs are then input into corresponding Encoders to obtain vector representation matrices of the sample protein molecule and the sample alternative molecule respectively, and the two vector representation matrices obtained are subjected to Graph Matching and Neighborhood Consensus as described above, while prediction of the sample activity value is obtained for the vector representation matrices of the sample protein molecule and the sample alternative molecule through an output layer of a neural network.
The server can determine whether the first training loss meets a first training target, the second training loss meets a second training target, and the third training loss meets a third training target, respectively. The training loss of the to-be-trained molecular binding model is determined to meet the training target when the first training loss, the second training loss and the third training loss all meet the corresponding training targets.
As an embodiment, the training target may be a preset target error value or a convergence state, without specific limitation.
After obtaining the first training loss, the second training loss and the third training loss, it can be determined whether the first training loss converges, whether the second training loss converges, and whether the third training loss converges, respectively, thereby determining whether the training loss of the to-be-trained molecular binding model meets the training target. In the event that the first training loss, the second training loss and the third training loss all converge, it is determined that the training loss of the to-be-trained molecular binding model meets the training target.
S503: during determining that the obtained training loss does not meet the training target, adjust model parameters of the to-be-trained molecular binding model.
If the obtained training loss does not meet the training target, it means that the predictive ability of the to-be-trained molecular binding model is not accurate enough and further training is needed and current model parameters need to be further adjusted, then the model parameters of the to-be-trained molecular binding model are adjusted.
As an embodiment, when in the first training loss, the second training loss and the third training loss, there is at least one training loss that does not converge, then it is determined that the obtained training loss does not meet the training target and the model parameters of the to-be-trained molecular binding model are adjusted.
S504: During determining that the obtained training loss meets the preset training target, output the molecular binding model as a trained molecular binding model.
If the obtained training loss meets the training target, it means that the predictive ability of the to-be-trained molecular binding model is already more accurate and current model parameters can no longer be adjusted, then the to-be-trained molecular binding model is output as the trained molecular binding model.
As an embodiment, the to-be-trained molecular binding model may be trained in a gradient manner, and the binding activity information may contain pIC50.
The training process in the embodiment of the present disclosure is implemented in an end-to-end format, without the need for artificial sub-problem partitioning, but is delivered entirely to the model to learn the mapping from original data to desired output directly, improving the intelligence for training the model and avoiding the problem that due to unpredictable circumstances in human involvement, the prediction accuracy of the trained model is low.
After obtaining the trained molecular structure model, the server can combine the feature extraction model, a target molecular structure model and the activity prediction model to predict the binding activity of the target protein molecule and the target alternative molecule after virtual binding. After obtaining the target protein molecule and the target alternative molecule, the server uses the feature extraction model to extract the feature of the target protein molecule and the target alternative molecule respectively, and obtains the target protein feature information and the target molecular feature information. Based on the target protein feature information and the target molecular feature information, the server uses the trained molecular structure model to obtain the binding activity feature information of the target protein molecule and the target alternative molecule after virtual binding. Based on the binding activity feature information, the server uses the activity prediction model to predict the activity value of the target protein molecule and the target alternative molecule after virtual binding.
The method for training molecular binding models provided in the embodiment of the present disclosure is described below, and please refer to
The server obtains the sample protein molecule and the sample alternative molecule, uses the feature extraction model to determine the adjacency matrix of the sample protein molecule and the adjacency matrix of the sample alternative molecule respectively, and uses the feature extraction model to perform feature extraction processing on the adjacency matrix of the sample protein molecule and the adjacency matrix of the sample alternative molecule respectively to obtain the protein feature information and the molecular feature information.
The server uses the to-be-trained molecular binding model to determine, based on the protein feature information and the molecular feature information, the binding activity feature information, the embedding feature information and the eutectic feature information between the sample protein molecule and the sample alternative molecule, based on graph matching, obtains the first error value between the embedding feature information and the embedding target, and based on neighborhood consensus, obtains the second error value between the eutectic feature information and the eutectic target. The server uses the activity prediction model to obtain, based on the binding activity feature information, a sample activity value, by training sample, obtains a reference activity value and determines the third error value between the sample activity value and the reference activity value.
The server determines the first training loss of the to-be-trained molecular binding model based on the first error value, determines the second training loss of the to-be-trained molecular binding model based on the second error value, and the third training loss of the to-be-trained molecular binding model based on the third error value.
When determining that there is a non-converging training loss in the various training losses, the server determines that the training loss of the to-be-trained molecular binding model does not converge and the parameters of the to-be-trained molecular binding model are adjusted. When determining that the training losses all converge, the server determines that the training loss of the to-be-trained molecular binding model converges, the to-be-trained molecular binding model is output, and the trained molecular binding model is obtained.
An embodiment of the present disclosure provides a method for screening molecules, and the trained molecular binding model trained and obtained by the method for training the molecular binding model described previously is used in performing molecular screening; please refer to
S701: Obtain target protein molecules and target alternative molecules.
In performing molecular screening, the target protein molecule and the target alternative molecule may be obtained first. There are various methods for obtaining the target protein molecule and the target alternative molecule, and as an embodiment, a target client may obtain the target protein molecule and the target alternative molecule in response to a molecular selection operation triggered by a target object. The target client sends the target protein molecule and the target alternative molecule to a server, and the server may receive the target protein molecule and the target alternative molecule sent by the target client.
As an embodiment, a protein molecule set and an alternative molecule set may be preset in the server, and the server may select a protein molecule in the protein molecule set as the target protein molecule and select an alternative molecule in the alternative molecule set as the target alternative molecule when performing molecular screening. The server, when performing selection in the protein molecule set or the alternative molecule set, may perform selection randomly or may, based on the attribute information of molecules contained in the set, perform selection according to a preset selection strategy, etc., without specific limitation.
S702: Use a feature extraction model to perform feature extraction processing on the target protein molecule and the target alternative molecule to obtain target protein feature information of the target protein molecule and target molecular feature information of the target alternative molecule.
S703: Use a trained molecular binding model to determine, based on the target protein feature information and the target molecular feature information, binding activity feature information between the target protein molecule and the target alternative molecule.
The trained molecular binding model may be trained by other devices and then sent to the server, and the server receives the trained molecular binding model sent by the other devices, or the trained molecular binding model may be a trained molecular binding model obtained by the server from the to-be-trained molecular binding model, without specific limitation.
S704: Use an activity prediction model to predict, based on the binding activity feature information, the activity value of the target protein molecule and the target alternative molecule after virtual binding.
After obtaining a virtual target compound, and virtual target embedding information and virtual target eutectic information corresponding to the virtual target compound, the server uses a target molecular structure model to obtain, based on the virtual target compound, the virtual target embedding information and the virtual target eutectic information, predicted activity information corresponding to the virtual target compound. The process is similar to the previously described process of using the to-be-trained molecular binding model to obtain, based on a sample compound, sample embedding information and sample eutectic information, sample activity information corresponding to the sample compound, and is not repeated here.
S705: Based on predicted marker activity information and preset molecular screening conditions, determine whether to select the target protein molecule and the target alternative molecule for combination.
After obtaining the predicted marker activity information, the server may determine, based on the predicted marker activity information and the preset molecular screening condition, whether to select the target protein molecule and the target alternative molecule for combination. The preset molecular screening condition may be a preset activity threshold, and the target protein molecule and the target alternative molecule are selected for combination when the predicted marker activity information is greater than or equal to the preset activity threshold. When the predicted marker activity information is less than the preset activity threshold, the target protein molecule and the target alternative molecule are not selected for combination.
The preset molecular screening condition may also be a preset ranking number, and after obtaining the predicted marker activity information of each group of target protein molecules and target alternative molecules, activity values contained in each predicted marker activity information are ranked in a descending order, and the target protein molecule and the target alternative molecule ranked before the preset ranking number are selected for combination.
An embodiment of the present disclosure provides an apparatus for training molecular binding models, and the apparatus may be a server as previously described, capable of achieving the functions corresponding to the previously described method for training molecular binding models. Please refer to
In an exemplary embodiment, the obtaining module 801 is specifically configured to:
In an exemplary embodiment, the obtaining module 801 is specifically configured to:
In an exemplary embodiment, the processing module 802 is specifically configured to:
In an exemplary embodiment, the processing module 802 is specifically configured to:
In an exemplary embodiment, the processing module 802 is specifically configured to:
In some embodiments, the processing module 802 is specifically configured to:
In an exemplary embodiment, the obtaining module 801 is further configured to:
In an exemplary embodiment, the processing module 802 is further configured to:
In an exemplary embodiment, the processing module 802 is specifically configured to:
An embodiment of the present disclosure provides an apparatus for screening molecules, and the apparatus may be a server as previously described, capable of achieving the functions corresponding to the previously described method for screening molecules. Please refer to
An embodiment of the present disclosure provides a computer device, the computer device 1000 being described below.
Please refer to
In one possible embodiment, the display panel 1041 may be configured in the form of a Liquid Crystal Display (LCD) or an Organic Light-Emitting Diode (OLED), etc.
The processor 1080 is configured to read a computer program and then execute a method defined by the computer program, for example, the processor 1080 reads a data storage program or file, etc., so as to run the data storage program on the computer device 1000 and display a corresponding interface on the display unit 1040. The processor 1080 may include one or more general purpose processors and may also include one or more Digital Signal Processors (DSP) for performing the relevant operations to implement the technical solutions provided by the embodiments of the present disclosure.
The memory 1020 generally includes an internal memory and an external memory, the internal memory may be a random access memory (RAM), a read-only memory (ROM), a high-speed cache (CACHE), etc. The external memory can be a hard disk, an optical disk, a USB disk, a floppy disk or a magnetic tape unit, etc. The memory 1020 is configured to store computer programs and other data, the computer program including applications corresponding to each client, etc., the other data including data generated after an operating system or application has been run, the data including system data (e.g., configuration parameters of the operating system) and user data. Program instructions in the embodiment of the present disclosure are stored in the memory 1020 and the processor 1080 executes the program instruction stored in the memory 1020 to implement any one of the method for training molecular binding models or the method for screening molecules as described in the preceding figures.
The display unit 1040 described above is configured to receive input digital information, character information or contact touch operations/non-contact gestures, and generate signal input related to user settings and function control, etc. of the computer device 1000. Specifically, in the embodiment of the present disclosure, the display unit 1040 may include a display panel 1041. The display panel 1041, such as a touch screen, may collect a touch operation of a user on or near the display panel (such as an operation of a user on the display panel 1041 or near the display panel 1041 by using any suitable object or accessory such as a finger or a stylus), and drive a corresponding connection apparatus according to a preset program.
In an exemplary embodiment, the display panel 1041 may include two parts: a touch detection apparatus and a touch controller. The touch detection apparatus detects a touch location of a player, detects a signal generated by a touch operation, and transmits the signal to the touch controller. The touch controller receives touch information from the touch detection apparatus, converts the touch information into touch point coordinates, and then sends the touch point coordinates to the processor 1080, and can receive and execute a command sent from the processor 1080.
The display panel 1041 may be implemented as display panels in various types of resistive, capacitive, infrared, or surface sound wave types. In addition to the display unit 1040, the computer device 1000 may also include an input unit 1030, the input unit 1030 may include a graphics input device 1031 and other input devices 1032, where the other input devices may include, but are not limited to, one or more of a physical keyboard, function keys (such as volume control buttons, switch buttons, etc.), a trackball, a mouse, an operating stick, and the like.
In addition to the above, the computer device 1000 may include a power supply 1090 for supplying power for other modules, an audio circuit 1060, a near-field communication module 1070 and an RF circuit 109. The computer device 1000 may also include one or more sensors 1050, such as an acceleration sensor, a photo sensor, a pressure sensor, etc. The audio circuit 1060 specifically includes a loudspeaker 1061 and a microphone 1062, etc. For example, the computer device 1000 may collect the user's voice through the microphone 1062 to perform corresponding operations, etc.
As an embodiment, the number of the processor 1080 may be one or more, and the processor 1080 and the memory 1020 may be coupled or relatively independent.
As an embodiment, the processor 1080 in
As an embodiment, the processor 1080 in
A person of ordinary skill in the art may understand that all or some steps for implementing the foregoing method embodiment may be completed by hardware related to a program instruction, the foregoing program may be stored in a computer-readable storage medium, and when being executed, the program performs steps including the foregoing method embodiment. The foregoing storage medium includes any medium that can store program codes, such as a removable storage device, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc, etc.
Alternatively, when the above integrated unit in the present disclosure is implemented in the form of a software function module and sold or used as an independent product, the integrated unit may also be stored in a computer-readable storage medium. Based on such an understanding, the technical solutions of the embodiments of the present disclosure essentially or the part contributing to the related art may be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions used to cause a computer device (which may be a personal computer, a server, or a network device) to execute all or a part of the method described in the embodiments of the present disclosure. The foregoing storage medium includes any medium that can store program codes, such as a removable storage device, a ROM, a RAM, a magnetic disk, or an optical disc, etc.
The term module (and other similar terms such as submodule, unit, subunit, etc.) in this disclosure may refer to a software module, a hardware module, or a combination thereof. A software module (e.g., computer program) may be developed using a computer programming language. A hardware module may be implemented using processing circuitry and/or memory. Each module can be implemented using one or more processors (or processors and memory). Likewise, a processor (or processors and memory) can be used to implement one or more modules. Moreover, each module can be part of an overall module that includes the functionalities of the module.
Obviously, a person skilled in the art can make various modifications and variations to the present disclosure without departing from the spirit and scope of the present disclosure. In this case, if the modifications and variations made to the present disclosure fall within the scope of the claims of the present disclosure and their equivalent technologies, the present disclosure is intended to include these modifications and variations.
| Number | Date | Country | Kind |
|---|---|---|---|
| 202111213797.X | Oct 2021 | CN | national |
This application is a continuation application of PCT Patent Application No. PCT/CN2022/116095, filed on Aug. 31, 2022, which claims priority to Chinese Patent Application No. 202111213797.X, filed on Oct. 19, 2021, the entire contents of all of which are incorporated by reference in the present disclosure.
| Number | Date | Country | |
|---|---|---|---|
| Parent | PCT/CN2022/116095 | Aug 2022 | US |
| Child | 18195744 | US |
