Candidate List Selection Model for Active Learning

Abstract

A method and apparatus for generating a selection model based on machine learning (ML) techniques is provided, the selection model being used to select a candidate list of compounds that require specific attribute-related validation. An iterative process or feedback loop for generating the selection model may include: receiving a prediction result list output from an attribute model for predicting whether a plurality of compounds are associated with a specific attribute and an attribute model score; retraining the selection model based on the attribute model score and / or the prediction result list; selecting a candidate list of compounds from a plurality of compounds associated with the prediction result list using the retrained selection model; sending the selected candidate list of compounds for specific attribute-related validation, wherein another ML technique is used to update the attribute model based on the validation; and repeating the receiving and retraining of the selection model until it is determined that the selection model has been effectively trained.

Description

Technical Field

[0001] The present application relates to devices, systems and methods for selecting candidate lists of compounds for active learning and model validation. Background Art

[0002] Informatics is the application of computer and information technology and resources to interpret data in one or more academic and / or scientific fields. Chemoinformatics and bioinformatics involve the application of computer and information technology and resources to interpret chemical and / or biological data. This can include solving and / or simulating processes and / or problems in the fields of chemistry and / or biology. For example, these computer and information technology and resources can convert data into information, and then convert information into knowledge for rapid creation of compounds and / or improved decision making in the fields of drug identification, discovery, and optimization (by way of example only and not limitation).

[0003] Machine learning techniques are computational methods that can be used to design complex analytical models and algorithms to solve complex problems, such as creating and predicting whether a compound has one or more features and / or properties. Although there are many ML techniques that can be used or selected for predicting whether a compound has a specific property or feature, there is usually a lack of training data to properly train the ML techniques to generate a properly trained ML model to predict whether a compound has a specific property, which is referred to as an attribute model in this article. If an ML technique is used to generate an attribute model based on insufficiently labeled training data, the resulting attribute model may not be able to reliably predict whether a compound has a specific property for most compounds.

[0004] Generating labeled training datasets for training ML techniques to generate accurate and reliable attribute models to predict whether a compound has a specific attribute is expensive, time consuming, and prone to human error. The complexity of this task increases exponentially as the number of attributes / features that need to be predicted increases, requiring the use of multiple attribute models to predict whether a compound has one or more of the multiple attributes and / or features. There is a need to improve the generation of labeled training datasets and select only those compounds with unknown associations with specific attributes to maximize the quality of the attribute models while minimizing the number of selected compounds. This will further improve training ML techniques for generating accurate and reliable attribute models to predict whether a compound has one or more specific attributes and help researchers, data scientists, engineers, and analysts make rapid progress in the field of drug identification, discovery, and optimization.

[0005] The embodiments described below are not limited to implementations that solve any or all of the disadvantages of the known approaches described above. Summary of the invention

[0006] The purpose of this disclosure is to introduce some concepts that will be further described in the following detailed description in a simplified form. This disclosure is not intended to illustrate the key features or essential features of the claimed subject matter, nor is it intended to be used to determine the scope of the claimed subject matter; variations and alternative features that contribute to the work of the present invention and / or are used to achieve substantially similar technical effects should be considered to fall within the scope of the present invention disclosed herein.

[0007] The present disclosure provides a method and apparatus for selecting a model, which predicts the best selection of a candidate list of compounds from a list of prediction results output by an ML model (e.g., an attribute model) for predicting whether a compound exhibits or has a specific attribute. The ML model is generated by training an ML technique based on a labeled training data set, and the labeled training data set is associated with a compound exhibiting a specific attribute / associated with a specific attribute. The selection model can be generated based on an iterative and semi-automatic / automatic method for training another ML technique, such as but not limited to a reinforcement learning (RL) technique for selecting the best candidate list of compounds from a list of prediction results. The RL technique is trained to generate a selection model that predicts the best selection of a candidate list of compounds for verification from a list of prediction results of an ML model (e.g., an attribute model), wherein the verification results from the candidate list of compounds can be used to enhance the ML model. The verification of the selected candidate list can include verifying the association of each compound in the selected candidate list with a specific attribute. The verification results can include an indication of each compound associated with a specific attribute; these can be used to generate a further labeled training data set to update or enhance the ML model. Therefore, the selection model helps to improve the training of the ML technique to generate an accurate and reliable attribute model.

[0008] The generated selection model can be applied when the labeled training data used to train the ML technique is insufficient to generate a sufficiently accurate property model for predicting whether a compound has a specific property. The selection model is configured to enhance the labeled training data set by selecting an optimal subset of compounds that should be validated with respect to a specific property, which, when validated, will update the property model and maximize or at least improve its predictive performance. For example, the validation results of the selected candidate list of compounds can be used to generate a further labeled training data set, which can be used to retrain the ML technique to generate an updated property model that maximizes or at least improves the predictive performance of the property model.

[0009] RL techniques can also be trained to generate selection models and / or validation models for predicting which validation method can be selected to validate the selected candidate list of compounds. For example, the selection model (validation model) can be further configured to predict whether and / or when to select a validation method based on computer analysis / simulation or a validation method based on laboratory experiments to validate the selected candidate list of compounds against a specific attribute. This can further minimize costs, such as but not limited to time, computing resources, and / or expenses associated with validating the candidate list of compounds.

[0010] In a first aspect, the present disclosure provides a computer-implemented method for generating a selection model based on ML techniques, the selection model being used to select a candidate list of compounds that need to be verified with a specific attribute. The method may use an iterative process / feedback loop, through which the selection model can be generated until it is determined that the selection model has been effectively trained. By way of example only, the process of each iteration of the feedback loop may include, but is not limited to: receiving a list of prediction results output from an attribute model for predicting whether a plurality of compounds are associated with a specific attribute and an attribute model score; training / retraining or updating the selection model based on the attribute model score and / or the prediction result list; selecting a candidate list of compounds from a plurality of compounds associated with the prediction result list using the retrained selection model; sending the selected candidate list of compounds for verification based on a specific attribute, wherein another ML technique is used to update the attribute model based on the verification; at least repeating the receiving and retraining of the selection model until it is determined that the selection model has been effectively trained. The process / loop may be repeated using an updated selection model until it is determined that the selection model has been effectively trained.

[0011] Preferably, the selection model can be generated by training an ML technique based on the attribute model score. As an option, the selection model can be generated by training an ML technique associated with reinforcement learning (e.g., an RL technique) based on the attribute model score and / or the list of predicted results. Preferably, the ML technique for generating and / or updating the attribute model is trained or updated based on a labeled training data set, which includes data representing a subset or set of multiple compounds and their known associations with specific attributes.

[0012] Preferably, retraining the selection model further comprises determining whether to retrain the selection model to select the candidate list of compounds based on the attribute model score and the previous attribute model score.

[0013] Preferably, the method further comprises: using the selection model, selecting a group of compounds from the prediction result list for the compound candidate list for verification; verifying the association of each compound in the selected compound candidate list with a specific attribute; and sending the verified candidate list of compounds to the ML technique for updating the attribute model based on the ML technique and the verified compound candidate list. For example, the verified candidate list of compounds can be sent to the ML technique for generating an attribute model based on the verified compound candidate list; the ML technique can be retrained based on the verified compound candidate list and the labeled training data set to update the attribute model.

[0014] Preferably, the ML technique generates the attribute model using a labeled training data set associated with a subset of multiple compounds related to a specific attribute, and wherein a validation candidate list of compounds or a validated compound candidate list (e.g., validation results involving the compound candidate list) is incorporated into the labeled training data set to update the attribute model by the ML technique that generates the attribute model.

[0015] Preferably, validating the selected candidate list of compounds further comprises validating the association of each compound in the candidate list of compounds with a specific attribute based on laboratory experiments or computer analysis.

[0016] Preferably, the laboratory experiment outputs a set of laboratory experiment validation results associated with the selected candidate list of compounds and the attributes, wherein the attribute model is updated based on the laboratory experiment validation results and the ML technique used to generate the attribute model. Preferably, the computer analysis outputs a set of computer analysis validation results associated with the selected candidate list of compounds and the specific attributes, wherein the attribute model is updated based on the computer analysis validation results and the ML technique used to generate the attribute model.

[0017] Preferably, the selection model is further trained to select a verification method for verifying the selected compound candidate list, wherein the verification method selected for verifying the selected compound candidate list further includes: selecting whether to perform a laboratory experiment or a computer analysis based on the specific attribute and the selected compound candidate list; in response to the selection, performing a laboratory experiment, wherein the laboratory experiment outputs a laboratory experiment verification result for estimating the association between each compound on the selected compound candidate list and the specific attribute, wherein the laboratory experiment verification result is used by the ML technology for generating the attribute model to update the attribute model; in response to the determination, performing a computer analysis, wherein the computer analysis outputs a computer analysis verification result for estimating the association between each compound on the selected compound candidate list and the specific attribute, wherein the computer analysis verification result is used by the ML technology for generating the attribute model to update the attribute model.

[0018] Preferably, the method also includes: receiving a prediction result list output from the attribute model for predicting whether multiple compounds are associated with a specific attribute and an attribute model score; determining whether to retrain the selection model to select a candidate list of compounds and a verification method based on the attribute model score and a previous attribute model score; and retraining the selection model based on the attribute model score and / or the prediction result list.

[0019] Preferably, when the verification method of performing laboratory experiments is selected and the number of iterations for retraining or updating the selected model is below a predetermined threshold, the method further comprises: penalizing the selected model during retraining; and selecting the verification method of performing computer analysis.

[0020] Preferably, when the validation method of performing laboratory experiments is selected and it is determined that the validation method of performing computer analysis will further improve the attribute model score, the method further includes: penalizing the selected model during retraining; and selecting the validation method of performing computer analysis.

[0021] Preferably, when a validation method of performing laboratory experiments is selected, and the selected candidate compound list has significant changes from a previously selected candidate compound list, the method further comprises: penalizing the selection model during retraining; and selecting a validation method of performing computer analysis.

[0022] Preferably, when a validation method that performs computer analysis is selected and it is determined that the computer analysis will improve the attribute model score of the attribute model based on a previous attribute model score calculated from a corresponding list of predicted results generated after each compound candidate list has been validated, the method further comprises: rewarding the selected model during retraining; and selecting a validation method that performs computer analysis.

[0023] Preferably, the prediction results list includes a predicted property score indicating each compound's association with a particular property.

[0024] Preferably, the predicted property score includes a certainty score, wherein compounds known to have a particular property are given a positive certainty score, compounds known not to have a particular property are given a negative certainty score, and other compounds are given an uncertainty score between positive and negative certainty scores.

[0025] Preferably, the certainty score is a percentage certainty score, wherein a positive certainty score is 100%, a negative certainty score is 0%, and the uncertainty score is between the positive and negative certainty scores.

[0026] Preferably, retraining the selection model further comprises: instructing the ML technology associated with the attribute model to restore the attribute model to the previous attribute model when the attribute model score does not reach the attribute model performance threshold compared to the corresponding previous attribute model score; instructing the ML technology to retain the updated attribute model on the previous attribute model when the attribute model score indicates that the attribute model performance threshold is met or exceeded compared to the corresponding previous attribute model score; and retraining the selection model to select a group of compounds from the corresponding list of predicted results based on the attribute model score; and repeating the first aspect, its modification or the steps as described herein until it is determined that the selection model has been trained.

[0027] Preferably, determining that the selection model is trained further comprises: comparing the retained attribute model scores with previously retained attribute model scores; and determining that the selection model has been effectively trained based on a plateau of the attribute model scores.

[0028] Preferably, the ML technique used to generate and / or update the selection model comprises at least one ML technique or a combination of ML techniques selected from: a recurrent neural network; a convolutional neural network; a reinforcement learning algorithm; and any neural network structure.

[0029] Preferably, the specific attributes include attributes or characteristics indicating the following: the compound docks with another compound to form a stable complex; the ligand docks with a target protein, wherein the compound is the ligand; the compound docks or binds to one or more target proteins; the compound has a specific solubility or solubility range; the compound has a specific toxicity; any other attribute or feature associated with the compound that can be simulated based on computer simulation and physical motion of atoms and molecules; any other attribute or feature associated with the compound that can be determined from an expert knowledge base; and any other attribute or feature associated with the compound that can be determined experimentally. The specific attributes may further include attributes, features and / or characteristics indicating the following: partition coefficient (e.g., LogP), partition coefficient (e.g., LogD), solubility, toxicity, drug-target interaction, drug-drug interaction, off-target drug effect, cell penetration, tissue penetration, metabolism, bioavailability, excretion, absorption, drug-protein binding, drug-lipid interaction, drug-deoxyribonucleic acid (DNA) / ribonucleic acid (RNA) interaction, metabolite prediction, tissue distribution and / or any other suitable attribute, feature and / or characteristic associated with the compound.

[0030] Preferably, the step of retraining the selection model further comprises retraining the selection model based on the attribute model scores and the prediction result list. As an option, retraining the selection model based on the prediction result list further comprises retraining the selection model based on one or more differences between a validation result associated with the validation of the candidate list of compounds and a current or previous prediction result list output from a current or previous attribute model. Preferably, wherein retraining the selection model using the prediction result list comprises retraining the selection model based on the structure of the compound, the attribute model is likely to incorrectly predict the structure of the compound.

[0031] Preferably, the ML technique for generating and / or updating the attribute model is trained or updated based on a labeled training data set, which includes data representing a subset or set of multiple compounds and their known associations with specific attributes. The ML technique for generating and / or updating the selection model is trained or updated based on the attribute model score and / or based on a list of predicted results output from the attribute model for multiple compounds. The attribute model score is a measure of the performance (e.g., predictive performance) of the attribute model after the attribute model has been trained when given a test set of labeled training data sets.

[0032] Preferably, the method for generating the selection model can be repeated until it is determined that the selection model has been effectively trained. In addition, the method may include further training the selection model by repeating the following steps, which steps include: receiving a list of prediction results; retraining or updating the selection model; selecting a candidate list of compounds; sending the candidate list for verification, wherein the verification result is used to update the attribute model; when repeating these method steps in the next iteration, the attribute model updated in the previous or current iteration will be used. Repeat these steps until it is determined that the selection model has been effectively trained, or the stopping criteria have been reached or met. For example, by analyzing the convergence or divergence of the attribute model scores of the attribute models updated in each iteration, it can be determined that the selection model is effectively trained. When the attribute model score does not substantially change to one or more previous and recent attribute model scores when the selected candidate list of compounds is subjected to computer analysis or laboratory experiments, the selection model can be considered to be effectively trained. In another example, the stopping criteria may include the maximum number of iterations that have been reached.

[0033] In a second aspect, the present disclosure provides a device for generating a selection model based on ML technology, wherein the selection model is used to select a candidate list of compounds that need to be verified with a specific attribute, and the device is configured to: receive a prediction result list output from an attribute model for predicting whether multiple compounds are associated with a specific attribute and an attribute model score; retrain or update the selection model based on the attribute model score and / or the prediction result list; use the retrained / updated selection model to select a candidate list of compounds from multiple compounds associated with the prediction result list; send the selected candidate list of compounds for verification based on the specific attribute, wherein another ML technology is used to update the attribute model based on the verification; at least repeat the reception and retraining of the selection model until it is determined that the selection model has been effectively trained.

[0034] In a second aspect, the present disclosure provides a device comprising a processor, a memory unit and a communication interface, wherein the processor is connected to the memory unit and the communication interface, wherein the processor, the communication interface and / or the memory unit are configured to implement a computer-implemented method according to the first aspect, its modification and / or as described herein.

[0035] In a third aspect, the present disclosure provides an ML model comprising data representing an ML selection model generated according to the first aspect, a modification thereof and / or a computer-implemented method as described herein.

[0036] In a fourth aspect, the present disclosure provides an ML selection model obtained or obtainable from a computer-implemented method according to the first aspect, a modification thereof and / or as described herein.

[0037] In a fifth aspect, the present disclosure provides a device comprising a processor, a memory unit and a communication interface, wherein the processor is connected to the memory unit and the communication interface, wherein the processor and the memory are configured to implement the ML selection model according to the third or fourth aspect, its modifications and / or as described herein.

[0038] In a sixth aspect, the present disclosure provides a computer-readable medium comprising data or instruction code representing an ML selection model generated based on training an ML technique according to the first aspect, its modification and / or the computer-implemented method as described herein, which, when executed on a processor, causes the processor to implement the ML selection model.

[0039] In a seventh aspect, the present disclosure provides a computer-readable medium comprising data or instruction code representing an ML selection model according to any of the first, third or fourth aspects, modifications thereof and / or as described herein, which, when executed on a processor, causes the processor to implement the ML selection model.

[0040] In an eighth aspect, the present disclosure provides an ML validation model obtained or obtainable from the first aspect, a modification thereof and / or a computer-implemented method as described herein.

[0041] In a ninth aspect, the present disclosure provides an ML selection and validation model obtained or obtainable from any of the first, third, fourth or eighth aspects, modifications thereof and / or a computer-implemented method as described herein.

[0042] In a tenth aspect, the present disclosure provides a computer-implemented method for updating an attribute model for predicting whether a compound is associated with a specific attribute, the method comprising: generating a result list of compounds using the attribute models of multiple compounds; selecting a candidate list of compounds from multiple compounds using a selection model according to any of the first aspect, the third aspect, the fourth aspect, or the eighth aspect, modifications thereof, and / or as described herein; receiving verification results of the candidate list of compounds; and updating the attribute model based on the verification results.

[0043] Preferably, the method further comprises repeating the steps of generating, selecting and receiving until it is determined that the attribute model is effectively trained.

[0044] Preferably, the method further comprises: verifying the association between each compound in the candidate compound list and a specific attribute, wherein the verification outputs a verification result, the verification result comprising data representing further labeled training data, the further labeled training data corresponding to the verified attribute association of each compound in the candidate compound list.

[0045] Preferably, wherein the property model is updated based on training a machine learning technique on a training data set of labels corresponding to a plurality of compounds and their association with a particular property.

[0046] In an eleventh aspect, the present disclosure provides a device comprising a processor, a memory unit, and a communication interface, wherein the processor is connected to the memory unit and the communication interface, wherein the processor, the communication interface and / or the memory unit are configured to implement a computer-implemented method according to the tenth aspect, its modification and / or as described herein.

[0047] In a twelfth aspect, the present disclosure provides a system for generating a selection model based on ML technology, the selection model being configured to select a candidate list of compounds for verification based on specific attributes, the system comprising: a selection module or device for selecting a candidate list of compounds according to any one of the first to ninth aspects, modifications thereof and / or as described herein; and an ML / attribute update module or device according to any one of the eleventh or twelfth aspects, modifications thereof and / or as described herein, the attribute update module or device being coupled to the selection module, wherein the attribute update module or device is configured to update the attribute model based on the selected candidate list of compounds.

[0048] In the thirteenth aspect, the present disclosure provides a system comprising: an attribute model generation module configured to generate an attribute model based on a labeled training data set, the labeled training data set comprising data representing compounds associated with a specific attribute; an ML verification module configured to generate a prediction result list output from the generated attribute model, for predicting whether multiple compounds are associated with a specific attribute and an attribute model score associated with the prediction; and a selection module configured according to any one of the first to ninth aspects, for selecting a compound candidate list for verification from the prediction result list, wherein the verification compound candidate list output is used to update the verification result of the attribute model.

[0049] Preferably, the property model generation module is further configured to receive a further labeled training data set based on the verification result of the compound candidate list, and update the property model by generating a property model based on the labeled training data set and the further labeled training data set.

[0050] The methods described herein may be performed by software in a machine-readable form on a tangible storage medium, for example in the form of a computer program comprising computer program code means adapted to perform all the steps of any method described herein when the program is run on a computer, and wherein the computer program may be implemented on a computer-readable medium. Examples of tangible (or non-transitory) storage media include disks, thumb drives, memory cards, etc., excluding propagating signals. The software may be adapted to be executed on a parallel processor or a serial processor, such that the method steps may be performed in any suitable order or simultaneously.

[0051] This application recognizes that firmware and software may be valuable, separately tradable commodities. It is intended to include software that controls or runs on "dumb" or standard hardware to perform a desired function. It is also intended to include software that "describes" or defines the configuration of hardware, such as HDL (Hardware Description Language) software used to design silicon chips or configure general-purpose programmable chips to perform a desired function.

[0052] As will be apparent to the skilled person, the preferred features may be combined as appropriate and may be combined with any aspect of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Embodiments of the present invention will be described by way of example with reference to the following drawings, in which:

[0054] Figure 1a is a schematic diagram showing an example training device / system for training a selection model to predict a candidate list of compounds according to the present invention;

[0055] Figure 1b It is shown that the Figure 1a A flowchart of an example training process for a training device;

[0056] Figure 1c is a schematic diagram showing another example training device for training a selection model to predict candidate list compounds according to the present invention, the device being used to update a property model to predict whether a compound has a specific property;

[0057] Figure 1d is a flowchart showing another example training process according to the present invention, the training process is used to Figure 1c training equipment;

[0058] Figure 2 is a table showing a list of example prediction results output from the property models of multiple compounds according to the present invention, for input to Figure 1b and 1c training process;

[0059] Figure 3 is a schematic diagram showing another exemplary selection device for selecting and verifying a candidate list of compounds according to the present invention; and

[0060] Figure 4 is a schematic diagram of a computing device according to the present invention.

[0061] The same reference numerals are used throughout the drawings to denote similar features. DETAILED DESCRIPTION

[0062] Embodiments of the present invention are described below by way of example only. These examples represent the best modes currently known to the applicant for putting the present invention into practice, but they are not the only implementations. This detailed description describes the functions of the examples and the order of steps for constructing and operating the examples. However, the same or equivalent functions and sequences may be implemented by different examples.

[0063] The inventors have advantageously developed a selection method / mechanism that generates a selection model for predicting the best selection of a candidate list of compounds from a machine learning (ML) model, such as a list of prediction results of an ML model, for example, by way of example only but not limitation, a property model for predicting whether a compound exhibits or has a specific property. The selection mechanism uses an iterative and semi-automatic / automatic method to train a reinforcement learning (RL) technique to generate a selection model for predicting the best selection of a candidate list of compounds. The selection model helps to enhance the training of the ML technique to generate an accurate and reliable property model for predicting whether a compound exhibits or has a specific property. The generated selection model and selection method / mechanism are particularly suitable when there is not enough labeled training data to train the ML technique to generate a property model for predicting whether a compound has a specific property. The selection model can be used to enhance the labeled training data set by selecting the best subset or candidate list of compounds for verification from the list of prediction results of the property model, which can further enhance the property model. For example, the verification results corresponding to the candidate list of compounds can be used to generate further labeled training data for updating the property model to maximize or at least improve its predictive performance. The selection mechanism can also learn to predict when to best verify the subset / candidate list based on a specific property through computer analysis or laboratory experiments.

[0064] Although the following description of the present invention only mentions, by way of example and not limitation, an attribute model and / or ML model for predicting whether one or more compounds are associated or have a specific attribute (e.g., whether one or more entities are associated with a certain relationship), those skilled in the art will appreciate that the present invention can be applied to other ML models for predicting whether an entity or input data has a specific relationship with another entity, or for classifying one or more entities and / or input data based on a specific relationship, etc. An entity may include one or more compounds, molecules, drugs, targets, proteins / genes, or other biological entities, etc.

[0065] A predictive attribute model (or ML model for predicting whether a compound exhibits or has a particular attribute) can be configured to receive a compound as input and output data representing a prediction of whether the compound has a particular attribute. For example, an attribute model can be configured to, by way of example only and not limitation, predict whether a compound will bind to a particular protein; or whether the compound is soluble in water; or predict whether the compound is toxic to the human body or a portion of the human body; or predict any other attribute of interest associated with a compound. However, labeled training data sets may only contain data for a few hundred to a few thousand compounds associated with a particular attribute. This data is insufficient to train an ML technique to generate an attribute model to predict whether a compound exhibits and / or has a particular attribute.

[0066] The quality of the attribute model can be improved by increasing the size of the labeled training data set. For example, multiple compounds with unknown associations with a specific attribute can be tested experimentally in the laboratory to measure whether they exhibit or are associated with a specific attribute. However, this is extremely expensive for most compounds. The inventors have developed a technique for limiting the number of compounds necessary to be tested in the laboratory while improving the quality of the attribute model. This can be achieved by using RL technology to iteratively learn how to select a candidate list of compounds from a list of prediction results for multiple compounds output from the attribute model, which most effectively improves the quality of the attribute model when retrained. RL technology can also iteratively learn to predict what kind of verification method should be used, for example, whether a candidate list of compounds associated with a specific attribute should be verified using computer analysis or laboratory experiments.

[0067] Assuming that the selection model has been trained to be suitable for selecting a candidate list of compounds and / or a verification method (e.g., computer analysis or laboratory experiments), whenever a plurality of compounds are presented to the attribute model, the selection model can be used to select a candidate list of compounds, most of which may not have been seen by the attribute model before (i.e., not part of the labeled training data set for the initial training attribute model). Typically, the selected candidate list may be more than the number of compounds generally sent to the laboratory for testing. Computer analysis, such as but not limited to computer simulations based on atomic and / or physical molecular interaction simulations (e.g., molecular dynamics RTM), can be used to verify the selected candidate list of compounds associated with a specific attribute. This assumes that a specific attribute can be simulated. The verification results of the candidate list are fed back to the attribute model (e.g., by enhancing the labeled training data set and retraining or updating the attribute model accordingly), and the attribute model can output another list of predicted results based on the or other multiple compounds. The selection model further predicts another candidate list for verification by simulation and feeds back to the attribute model. These steps can be repeated until the selection model predicts a candidate list suitable for laboratory testing to further improve the quality of the attribute model. After laboratory testing, the laboratory results of the validated compound candidate list can be fed back into the attribute model (e.g., the laboratory results are used to further enhance the labeled training data set and the attribute model is retrained accordingly). These steps can be repeated to select compound candidate lists for computer analysis and / or laboratory experiments until the attribute model is deemed to have been appropriately trained.

[0068] A compound (also referred to as one or more molecules) may include or represent a chemical or biological substance composed of one or more molecules (or molecular entities) composed of atoms from one or more chemical elements (or more than one chemical element) held together by chemical bonds. Exemplary compounds used herein may include, by way of example only, but not limitation, molecules held together by covalent bonds, ionic compounds held together by ionic bonds, intermetallic compounds held together by metallic bonds, certain complexes held together by coordinate covalent bonds, pharmaceutical compounds, biological compounds, biomolecules, biochemical compounds, one or more proteins or protein compounds, one or more amino acids, lipids or lipid compounds, carbohydrates or complex carbohydrates, nucleic acids, deoxyribonucleic acid (DNA), DNA molecules, ribonucleic acid (RNA), RNA molecules, and / or any other organization or structure of molecules or molecular entities composed of atoms from one or more chemical elements and combinations thereof.

[0069] Each compound has or exhibits one or more properties, features or characteristics, or a combination thereof, which may determine the usefulness of the compound for a given application. The properties or associated attributes of a compound may include or represent data that indicates or indicates a specific behavior of the compound when the compound undergoes a reaction. For example, a compound may be associated with or exhibit one or more features or attributes, which may include, by way of example only, but are not limited to, one or more features or attributes from the following group: an indication that the compound docks with another compound to form a stable complex; an indication that a ligand docks with a target protein, wherein the compound is a ligand; an indication that the compound docks or binds to one or more target proteins; an indication that the compound has a specific solubility or solubility range; an indication that the compound has a specific electrical property; an indication that the compound has toxicity or a toxicity range; any other indication of a property or characteristic associated with the compound that can be simulated using computer simulation based on the physical motion of atoms and molecules; any other indication of a property or characteristic associated with the compound that can be experimentally or measured. Further examples of one or more compound attributes, features, or properties may include, by way of example only, but not limitation, one or more of: Log, LogD, solubility, toxicity, drug-target interaction, drug-drug interaction, off-target drug effect, cell penetration, tissue penetration, metabolism, bioavailability, excretion, absorption, drug-protein binding, drug-lipid interaction, drug-DNA / RNA interaction, metabolite prediction, tissue distribution, and / or any other suitable attribute, feature, and / or property associated with the compound.

[0070] The properties of a given compound may include data representing or indicating a particular behavior / characteristic / property of the compound when the compound undergoes a reaction, which data representing or indicating the properties of the compound may include, by way of example only, but is not limited to, any continuous or discrete value / score and / or range of values / scores, a series of values / scores, a string of characters, or any other data representing the property. For example, a property may be associated with, assigned by, represented by, or based on, one or more continuous property values / scores (e.g., non-binary values), one or more discrete property values / scores (e.g., binary values), one or more ranges of continuous property values / scores, one or more ranges of discrete property values / scores, a series of property values / scores, one or more property value strings, or any other suitable data representation of property values / scores representing a property. The property values / scores may be based on measured or simulated data associated with the reaction and / or the particular property.

[0071] When a compound experiences a reaction associated with a specific attribute, a property value / score may be assigned to the compound, the property value / score including data indicating whether they are associated with a specific attribute. By way of example only, the property value / score may be determined or based on laboratory measurements and / or computer simulation values / scores, but is not limited thereto. The property value / score assigned to a compound gives an indication of whether the compound is associated with or exhibits a specific attribute. For example, when a compound experiences a reaction associated with a specific attribute, a property value / score may be assigned to the compound, depending on whether the compound exhibits a specific attribute. When the property value / score associated with a compound, by way of example only but not limited thereto, is higher or lower than a threshold property value / score representing the attribute, within or near a region representing the value of the attribute, etc., it can be said that the compound exhibits a specific attribute.

[0072] According to the present invention, the property model generated for predicting whether a compound has one or more properties can be generated using one or more ML techniques or a combination thereof. ML techniques may include or represent one or more computational methods or a combination of computational methods that can be used to generate analytical models and algorithms that help solve complex problems, such as, by way of example only but not limitation, the prediction and analysis of complex processes and / or compounds. ML techniques can be used to generate property models for drug discovery, identification and / or optimization in the fields of informatics, cheminformatics and / or bioinformatics.

[0073] For example, a labeled training data set can be used to train ML techniques to generate an ML model (e.g., a property model for predicting whether a compound has a specific property). The labeled training data set includes data representing one or more compounds, each of which can be labeled with data representing known property values / scores or tags associated with the compound and a specific property, and the data set can be used by the ML technology to generate a property model. Therefore, once the ML technology has trained an ML model based on such a labeled training data set associated with a specific property, the resulting trained ML model can be referred to as a property model. The property model can therefore predict or classify whether an input compound exhibits a specific property. The property model can output data representing a property value / score, and the property value / score represents the association of the input compound with a specific property. The data representing the property value / score output by the property model can be referred to as a property prediction value / score here. Data representing one or more compounds can be input to the property model, and the property model can output a property prediction value / score, and the property prediction value / score includes data representing one or more corresponding property values / scores, and the property value / score indicates whether one or more input compounds are associated with or exhibit the specific property modeled by the property model.

[0074] Examples of ML techniques that can be used to generate an ML model or property model for predicting whether a compound has a particular property may include, by way of example only and not limitation, at least one of the following ML techniques or a combination of ML techniques: a recurrent neural network; a convolutional neural network; a reinforcement learning algorithm; and any other neural network structure configured to predict whether a compound has a particular property.

[0075] Further examples of ML techniques that may be used according to the invention described herein may include or be based on, by way of example only and not limitation, any ML technique or algorithm / method that may be trained or adapted to generate one or more candidate compounds based on, by way of example only and not limitation: an initial compound, a list of desired properties of a candidate compound, and / or a set of rules for modifying a compound. The ML technique or algorithm / method may include one or more of supervised ML techniques, semi-supervised ML techniques, unsupervised ML techniques, linear and / or nonlinear ML techniques, ML techniques associated with classification, ML techniques associated with regression, etc., and / or combinations thereof. Some examples of ML techniques may include or be based on, by way of example only and not limitation, one or more of active learning, multi-task learning, transfer learning, neural message parsing, one-shot learning, dimensionality reduction, decision tree learning, association rule learning, similarity learning, data mining algorithms / methods, artificial neural networks (NN), deep neural networks, deep learning, deep learning artificial neural networks, inductive logic programming, support vector machines (SVM), sparse dictionary learning, clustering, Bayesian networks, representation learning, similarity and metric learning, sparse dictionary learning, genetic algorithms, rule-based machine learning, learning classifier systems, and / or one or more combinations thereof, and the like.

[0076] Some examples of supervised ML techniques may include or be based on, by way of example only but not limitation, ANNs, DNNs, association rule learning algorithms, a priori algorithms, case-based reasoning, Gaussian process regression, grouping methods for data processing (GMDH), inductive logic programming, instance-based learning, lazy learning, learning automata, learning vector quantization, logical model trees, minimum message length (decision trees, decision diagrams, etc.), XGBOOST, gradient boosting machines, nearest neighbor algorithms, analogy modeling, probably approximately correct learning (PAC), downward fluctuation rules, knowledge acquisition methods, symbolic machine learning algorithms, support vector machines, random forests, classifier ensembles, bootstrap aggregation (BAGGING), boosting (meta-algorithms), ordered classification, informative fuzzy networks (IFNs), conditional random fields, analysis of variance, quadratic classifiers, k-nearest neighbors, boosting, sprinting, Bayesian networks, naive Bayes, hidden Markov models (HMMs), hierarchical hidden Markov models (HHMMs), and any other ML technique or ML task capable of inferring functions or generating models from labeled and / or unlabeled training data, etc.

[0077] Some examples of unsupervised ML techniques may include or be based on, by way of example only and without limitation, the expectation maximization (EM) algorithm, vector quantization, generating topographic maps, the information bottleneck (IB) method, and any other ML technique or ML task capable of inferring functions to describe hidden structures and / or generating models from unlabeled data and / or from labels in a labeled training dataset that is ignored, etc. Some examples of semi-supervised ML techniques may include or be based on, by way of example only and without limitation, active learning, generative models, low-density separation, graph-based methods, joint training, transformation, or any other ML technique, task, or class of unsupervised ML techniques capable of training with unlabeled datasets and / or labeled datasets, etc.

[0078] Some examples of artificial neural network (ANN) ML techniques may include or be based on, by way of example only and without limitation, one or more of ANN, feedforward neural network, recurrent neural network (RNN), convolutional neural network (CNN), autoencoder neural network, extreme learning machine, logic learning machine, self-organizing map and other ANN ML techniques or connectionist systems / computing systems inspired by biological neural networks that make up animal brains. Some examples of deep learning ML techniques may include or be based on, by way of example only and without limitation, one or more of deep belief network, deep Boltzmann machine, DNN, deep CNN, deep RNN, hierarchical temporal memory, deep Boltzmann machine (DBM), stacked autoencoder and / or any other ML techniques.

[0079] Although a set of selection and / or verification rules can be derived to select a candidate list of compounds and / or select a verification method for verifying a candidate list of compounds as described herein, a selection model can be generated based on training reinforcement learning techniques. The selection model is used to predict a candidate list of compounds suitable for verification of a specific attribute. Therefore, instead of using a set of selection rules to select a suitable candidate list of compounds that is uncertain about the attribute model, such selection can be performed by training the RL technique over a certain period of time. Once the RL technique has learned to select a candidate list of compounds for enhancing the attribute model, the generated selection model can be used to train an attribute model that is used to predict whether a compound exhibits or has an attribute different from a specific attribute. This is because the selection model does not depend on the attribute type predicted by each attribute model modeling.

[0080] In order to maximize the quality of the selection and generate a selection model, RL techniques can be trained to learn what compounds to select from the result prediction list. The quality of the selection can be maximized when the selected compound candidate list is the best compound selected from a specific result prediction list, so that the quality of the resulting updated attribute model can be maximized when validated with respect to a specific attribute. RL techniques can be used to iteratively train a selection model that is robust enough to select the most suitable or best compound candidate list from the result prediction list for validation based on a specific attribute.

[0081] The RL techniques described herein may be based on, by way of example only and not limitation, at least one ML technique or combination of ML techniques from: a recurrent neural network; a convolutional neural network; a reinforcement learning algorithm; any other neural network structure suitable for training RL techniques to generate a selection model as described herein; and / or any other ML or RL structure suitable for training RL techniques to generate a selection model as described herein.

[0082] Figure 1a 1 is a schematic diagram showing an example training device 100, which is used to generate a selection model based on RL technology, and the selection model is used to select a candidate list of compounds that need to be verified with a specific attribute from a prediction result list. The prediction result list can be output from a property model for predicting whether a plurality of compounds are associated with a specific attribute. For example, the prediction result list can include a plurality of compounds, each of which is mapped to a corresponding property prediction value / score output / estimated by the property model. The training device 100 performs an iterative feedback loop for training a selection model based on RL technology during training or updating of the property model. The property model is used to predict whether a compound is related to a specific attribute or displays or has a specific attribute. The training device 100 includes a property model generation (MLG) device 102, a model evaluation (ME) device 104, a selection mechanism (SM) 106, and a feedback path 107. The MLG device 102 is configured to train ML technology to generate or update a property model based on a first set of labeled training data sets. The generated / updated or trained property model can be input to the ME device 104. The ME device 104 is configured to test and evaluate the attribute model based on compounds from the second set of labeled training data sets, wherein the first set and the second set of labeled training data sets are different. The SM 106 is configured to generate a selection model based on training the RL technique using the attribute model score. The SM 106 feeds back a further labeled training data set via a feedback path 107 based on the selected candidate list of compounds, and verifies it against the second set of labeled training data sets to form a further labeled training data set. The attribute model is updated based on the ML technique to update the attribute model for further iterations of training.

[0083] Specifically, initially in the first iteration, the MLG device 102 receives a labeled training data set {T i} j , where N is the number of training data elements (e.g., in the range of 1000 or more), where the i-th training data element includes i and its known association with specific attributes. The MLG device 102 can use the first set of labeled training data sets {T i} j 1 To train the ML technique (which may be predetermined) to generate the attribute model M for the jth iteration j . Attribute Model M j Predict whether the input compound C1 has a specific attribute. The second set of labeled training data sets {T i} j 2 , where N = L + K, is used to train the RL technique to generate the selection model.

[0084] The ME device 104 receives the generated attribute model M j , multiple compounds {C l} j Input to the attribute model M j , where 1<=l<=L, L is the number of the plurality of compounds. In this case, the plurality of compounds {C l} j is the second set of labeled training data set {T i} j 2 Associated compounds. Property model M j Therefore, the output prediction result list {R l} j , where the l-th prediction result R for the j-th iteration l,j The data may include, by way of example only and not limitation, data representing compound C1 and the prediction score P for the jth iteration l,j . Prediction score P l,j M stands for the attribute model j The predicted value indicates that compound C1 is associated with a specific attribute. l} j Predict multiple compounds {C l} j Whether each of has a specific attribute.

[0085] The ME device 104 also based on the prediction result list {Rl} j and / or a second set of labeled training data sets {T i} j 2 To calculate the estimated attribute model M j The quality of the attribute model score S j . Attribute model score S j It can be calculated based on model performance statistics, which can be obtained from the labeled training dataset {T i} j , the second labeled training dataset {T i} j 2 and / or the received prediction result list {R l} j 200 to estimate. Model performance statistics may include or represent the training data set based on the label {T i} j , the second labeled training dataset {T i} j 2 and / or the received prediction result list {R l} j An indication of the performance of the attribute model of 200. The model performance statistics of the attribute model can be based on, by way of example but not limitation, one or more of: a positive predictive value or precision of the attribute model; a sensitivity, true prediction rate, or recall rate of the attribute model; a receiver operating characteristic (ROC) curve plot associated with the attribute model; an area under the precision and / or recall ROC curve associated with the attribute model; any other function related to the precision and / or recall rate of the attribute model; and an attribute model score S used to generate an attribute model score indicating the performance of the attribute model. j Any other model performance statistics.

[0086] The RL technique of the selection mechanism 106 may be taught to predict the list of results {R l} j Which compounds may be the best choices for validation, and thus generate a selection model. Initially, the selection mechanism 106 may only initialize the selection model, which still needs to be trained by RL technology, so it can initially be predicted from the result list {R l} j A set of "random" compounds is selected as the candidate list of selected compounds. i} j 2 Verify the selected compound candidate list to form a candidate list labeled training dataset {T k} j This is because the resulting prediction list {Rl} j The attribute model M j Using the training dataset from the second set of labels {T i} j 2 The compound {C l} generated, the second set of labeled training data sets {T i} j 2 The portion of the candidate list corresponding to the selected compound can be fed back through the feedback path 107 as a training data set {T k} j Input to MLG 102 for retraining / updating the attribute model M in the next iteration j The training device 100 then proceeds to perform the next iteration (eg, j=j+1).

[0087] In the second iteration (eg, j=2), the MLG device 102 may use the first set of labeled training data sets {T i} j 1 and a candidate list labeled training data set {T k} j To retrain the ML technique (which may be predetermined) to generate the attribute model M for the jth iteration j The ME device 104 receives the generated attribute model M from the MLG device 102 j , and input the second set of labeled training data sets {T i} j 2 Multiple compounds associated with l} j . Attribute Model M j Therefore, the output prediction result list {R l} j , where the l-th prediction result Rl,j for the j-th iteration may include, by way of example only but not limitation, data representing compound C1 and the prediction score P for the j-th iteration l,j .

[0088] The ME device 102 also further includes a prediction result list {R l} j and / or a second set of labeled training data sets {T i} j 2 Calculate the estimated attribute model M j Another attribute model score S of quality and / or performancej The attribute model scores {S k} from the previous iteration (e.g., k = j - 1) where 1 <= k < j can be compared with the attribute model scores S j of the current iteration. The attribute model scores {S k} where 1 <= k <= j and the prediction result list {R l} j can be fed into the selection mechanism 106. If the selection mechanism 106 determines that the quality / accuracy of the performance of the attribute model M k has improved based on the attribute model scores {S j}, it is fed back to the RL technique as a reward. The selection model associated with the RL technique can be updated / retrained based on the reward. Then, the selection mechanism 106 can use the updated selection model to select another set of compounds from the result prediction list {R l} j as a list of compound candidates for validation. The selected list of compound candidates is verified against the second set of labeled training data set {T i} j 2 to form a training data set {T k} j labeled with the candidate list. This is because the result prediction list {R l} j is generated by the attribute model M j using the compounds {Cl} from the second set of labeled training data set {T i} j 2 . The part of the second set of labeled training data set {T i} j 2 corresponding to the selected list of compound candidates can be fed back through the feedback path 107 as a training data set {T k} j input to the MLG 102 for retraining / updating the attribute model M j in the next iteration. The training device 100 then continues to perform the next iteration (e.g., j = j + 1).

[0089] However, if the selection mechanism 106 determines that the comparison of two or more attribute model scores {S k} where 1 <= k <= j finds that the quality / accuracy of the performance of the attribute model M j has not improved, it is fed back to the RL technique as a penalty. The selection model associated with the RL technique can be updated / retrained based on the penalty. Given the attribute model M jThe performance of the model M has deteriorated, it may be restored to the previous performance has not deteriorated property model M j-1 The selection model can then be used to predict the list {R l} j Another group of compounds is selected as a candidate list of compounds for verification. i} j 2 Verify the selected compound candidate list to form a candidate list labeled training dataset {T k} j This is because the resulting prediction list {R l} j The attribute model M j Using the training dataset from the second set of labels {T i} j 2 The compound {C l} generated, the second set of labeled training data sets {T i} j 2 The portion of the candidate list corresponding to the selected compound can be fed back through the feedback path 107 as a training data set {T k} j Input to MLG 102 for retraining / updating the attribute model M in the next iteration j The training device 100 then proceeds to perform the next iteration (eg, j=j+1).

[0090] The device 100 repeats the iterative feedback process multiple times until, for example only but not limited to, there are no more compounds that need to be selected for the candidate compound list; and / or the attribute model score {S k}1<=k<=j indicates that the performance of the ML technique has stabilized; or some other condition or criterion indicating that the selection model has been effectively trained and can be used to select further candidate lists of compounds.

[0091] Figure 1b is shown for generating reference Figure 1a Flowchart of an example training process 110 for a selection model is described. Assume that the RL technique is trained to generate a selection model for predicting a candidate list of compounds that will enhance further training of an attribute model. The attribute model is used to predict whether a compound is associated with a specific attribute. The training process 110 can be implemented by the selection mechanism 106 and / or the training device 100. The training process 110 can be as follows:

[0092] In step 112, a prediction result list is received from the attribute model for predicting whether a plurality of compounds are associated with a specific attribute and an attribute model score S j Associated. In step 114, based on the attribute model score, the previous attribute model score and / or the predicted result list, it is determined whether to retrain the selection model. If it is determined to retrain the selection model (e.g., "yes"), proceed to step 115. In step 115, the RL technology is retrained based on the attribute model score of the current iteration to generate an updated selection model. For example, if the attribute model score is improved compared to the previous iteration, the RL technology will be rewarded during the retraining, which will have a positive impact on the selection model. That is, the reward enhances the RL technology to select compounds that can improve the performance of the attribute model. If the attribute model score is not improved compared to the previous iteration, the RL technology will be punished during the retraining, which will have a negative impact on the selection model, that is, the RL technology learns not to make poor choices. In step 116, the selection model that may have been updated in step 115 is used to select a candidate list of compounds based on the predicted result list. In step 118, the selected candidate list of compounds is sent or fed back in the form of a further labeled training data set for the ML technology to update the attribute model. Steps 112 - 118 may be repeated until it is determined that the selection model has been effectively or sufficiently trained.

[0093] The training apparatus 100 and / or training process 110 can be used to initially guide the generation of a selection model using only a labeled training data set. However, the labeled training data set can be further enhanced using computer analysis / simulation and / or laboratory experiments to validate the selected unknown compounds associated with a particular attribute. This additional validation can also be used in the feedback path 107 to enhance the selection model and the attribute model M. j .

[0094] Then, the attribute model M can be further trained as described below j , where multiple compounds (most of which have not been seen before by the attribute model) can be input into the attribute model to generate a list of prediction results {R l} j, wherein the generated selection model can be used to select a candidate list of compounds for validation through computer analysis / simulation and / or laboratory experiments on specific attributes. As described above, the validation results can be used to generate further labeled training data, which can be used to further update the attribute model, thereby iteratively further improving the attribute model. In this process, the selection model can also be further trained by modifying the above-mentioned training device 100 to validate each selected candidate list of compounds based on specific attributes using computer analysis / simulation and / or in rare cases using laboratory experiments. Attribute model scores can be calculated to allow RL techniques to reward or penalize the selection model during retraining.

[0095] Figure 1c is based on Figure 1a Another schematic diagram of another training device 120 is shown, which is configured to select a candidate list of unknown compounds for verification related to a specific attribute to further enhance the training of the RL technique for generating the selection model, while also enhancing the training for generating the ML model M j The ML technology is used for training. In this example, the ML model M j The training device 120 includes an MLG device 102, an ME device 104, and an SM 106, wherein the SM 106 includes a selection model 106a and a verification device 106b, which is used to verify the candidate list of compounds associated with the specific attribute selected from the selection model. The verified candidate list of selected compounds is fed back in the form of a further labeled training data set {Tk}j. Based on the further labeled training data set {T k} j To update the attribute model for further iterations of training.

[0096] Specifically, initially in the first iteration, the MLG device 102 receives a labeled training data set {T i} j , where N is the number of training data elements (e.g., in the range of 1000 or more), where the i-th training data element includes i and its known association with specific attributes. The MLG device 102 can use the labeled training data set {T i} j To train the ML technique (which may be predetermined) to generate the attribute model M for the jth iteration j . Attribute Model M j is generated / updated to predict the input compound C l Whether it has specific attributes.

[0097] Once the ML technique has been trained appropriately to generate or update the attribute model M j , attribute model M j It can be sent to the ME device 104. The ME device 104 receives the generated attribute model M j , and multiple compounds {C l} j Input to the attribute model M j , where 1<=l<=L, and L is the number of the plurality of compounds. In this case, the plurality of compounds {C l} j Can include training data from labeled i} j but also includes a group of compounds with unknown associations or relationships with specific attributes, which may be larger than the set of unknown compounds from the labeled training data set {T i} j The attribute model Mj therefore outputs a list of predicted results {R l} j , where the l-th prediction result for the j-th iteration {R l} j It may include, by way of example only and not limitation, compounds C l The data and the attribute prediction score P for the jth iteration l,j . Attribute prediction score P l,j is the property model for compound C l M associated with a specific attribute j Predicted value. Prediction result list {R l} j Predict multiple compounds {C l} j Does each of them have a specific attribute? For example, the prediction result list {R l} j may include being input to the attribute model M j Multiple compounds {C l} j , where each compound C l can be mapped or assigned corresponding attribute prediction values / scores, which can be represented by the attribute model M j For each compound C l Output / Estimate.

[0098] The ME device 104 also based on the prediction result list {R l} j and / or labeled training dataset {T i} jTo calculate the attribute model score S j , the attribute model score S j Estimate the quality of the attribute model Mj. Attribute model score S j It can be calculated based on model performance statistics, which can be obtained from the labeled training dataset {T i} j and / or the received prediction result list {R l} j 200 to estimate. Model performance statistics may include or represent the training data set based on the label {T i} j and / or the received prediction result list {R l} j An indication of the performance of the attribute model of 200. The model performance statistics of the attribute model can be based on, by way of example but not limitation, one or more of: a positive predictive value or precision of the attribute model; a sensitivity, true prediction rate, or recall rate of the attribute model; a receiver operating characteristic (ROC) curve plot associated with the attribute model; an area under the precision and / or recall ROC curve associated with the attribute model; any other function related to the precision and / or recall rate of the attribute model; and an attribute model score S used to generate an attribute model score indicating the performance of the attribute model. j Any other model performance statistics.

[0099] The RL techniques of SM 106 can be taught to predict a list of outcomes {R l} j Which compounds in the list may be the best choices for validation, thereby generating a selection model. Initially, SM 106 may only initialize the selection model 106a, which still needs to be trained by RL technology, so that the result prediction list {R l} j A set of "random" compounds is selected as the candidate list of selected compounds. Optionally, SM 106 may have used feedback path 107 and the first and second sets of labeled training data sets to select a set of "random" compounds as the candidate list of selected compounds. Figure 1a and 1b The training apparatus 100 and process 110 described in the above guide the selection model 106a. In any case, the selection model 106a selects the candidate compound list for verification. Assuming that one or more selected candidate compound lists may have unknown associations or relationships with specific attributes, the selected candidate compound list is verified in the verification device 106b with respect to the specific attribute.

[0100] The verification device 106b can use a labeled training data set, computer analysis / simulation, and / or laboratory experiment (or multiple experiments) to verify one or more selected candidate lists of compounds based on a specific attribute to establish an association between each compound and a specific attribute. The verification device 106b outputs a labeled training data set of candidate lists {T k} j , which provides an indication of the association of each of the candidate lists of compounds with a particular attribute.

[0101] Assume that the prediction result list {R l} j The attribute model M j Using compound {C l} generated, the compound may include a labeled training dataset {T i} j The selected candidate list of compounds may include a plurality of compounds with known associations with a specific attribute and a plurality of compounds with unknown associations with a specific attribute from a labeled training dataset {T i} j One or more of the plurality of compounds and / or one or more of the plurality of compounds having an unknown association with a particular property.

[0102] The verification device 106b can detect whether one or more selected candidate compound lists include compounds from the labeled training data set {T i} j If from the labeled training data set {T i} j If one or more of the multiple compounds are detected in the candidate list, these compounds may not need to be verified using computer analysis / simulation. i} j The corresponding part of the training dataset is simply added to the candidate list {T k} j If one or more of the multiple compounds in the candidate list have unknown associations with a specific attribute, i.e., they do not appear in the model used to train the attribute model M, j Once the labeled training data set is obtained, the compounds can then be validated based on specific properties using computer analysis / simulation and / or sometimes or when necessary, laboratory experiments.

[0103] The validation results for these compounds can be used to generate additional portions of labeled training data elements to include in the candidate list labeled training data set {T k} jThen, based on the selected candidate list of compounds, the verification device 106b sends the training data set {T k} j Forwarded to MLG 102 as a training dataset labeled with candidate lists {T k} j Input, used to retrain / update the attribute model M in the next iteration j . Labeled training dataset {T k} j This candidate list can be merged into the original labeled training dataset {T i} j It can be used to retrain the attribute model M j The training device 100 then proceeds to perform the next iteration (eg, j=j+1).

[0104] In the second iteration (eg, j=2), the MLG device 102 may use the previously labeled training dataset {T i} j and the training dataset labeled with the previous candidate list {T k} j The enhanced labeled training dataset {T i} j To retrain or update the ML technique (which may be predetermined) to generate an updated attribute model M for the jth iteration j The ME device 104 receives the generated / updated attribute model M from the MLG device 102. j , and input multiple compounds {C l} j . Multiple compounds {C l} j It can be different from the previous iteration or the same. j Therefore, the prediction result list {R l} j , where the l-th prediction result R for the j-th iteration l,j It may include, by way of example only and not limitation, compounds C l data and the prediction score P for the jth iteration l,j .

[0105] The ME device 102 also further includes a prediction result list {R l} j and / or enhanced labeled training datasets {T i} j To calculate the property model M used to estimate jAnother attribute model score S of the quality and / or performance j The attribute model scores {S k} from the previous iteration (e.g., k = j - 1) where 1 <= k < j can be compared with the attribute model score S of the current iteration j The attribute model scores {S k} where 1 <= k <= j and the prediction result list {R l} j can be fed into SM106. If SM106 determines that the performance of the attribute model M k has improved in terms of quality / accuracy based on the attribute model scores {S j}, it is fed back as a reward to the RL technique. The selection model 106a associated with the RL technique can be updated / retrained based on this reward. Then, the selection mechanism 106 can use the updated selection model 106a to select another set of compounds from the result prediction list {R l} j as a list of compound candidates for verification by the verification device 106b. The selected list of compound candidates is verified using the verification device 106b to form a training data set {T k} j labeled with candidate lists in a manner similar to that described in the previous iteration. The training data set {T k} j labeled with candidate lists can be fed back via the feedback path 108 to be input into the MLG 102 as a training data set {T k} j for retraining / updating the attribute model M j in the next iteration. The training device 100 then continues to perform the next iteration (e.g., j = j + 1).

[0106] However, if the selection mechanism 106 determines that the comparison of two or more attribute model scores {S k} where 1 <= k <= j finds that the quality / accuracy of the performance of the attribute model M j has not improved, it is fed back as a penalty to the RL technique. The selection model 106a associated with the RL technique can be updated / retrained based on the penalty. Given that the performance of the attribute model M j has deteriorated, it may revert to the attribute model M j-1 whose performance had not deteriorated previously. Then, the selection model 106a can be used to select from the result prediction list {R l} jAnother group of compounds is selected as a candidate list of compounds to be verified by the verification device 106b. The selected candidate list of compounds is verified by the verification device 106b to form a training data set {T k} j The candidate list labeled training dataset {T k} j The training data set {T k} j The input is sent to the MLG 102 for retraining / updating the attribute model Mj in the next iteration. The training device 100 then continues to execute the next iteration (eg, j=j+1).

[0107] The device 120 repeats the iterative feedback process multiple times until, by way of example only and not limitation, there are no more compounds that need to be selected for the candidate compound list; there are no more compounds that have unknown relationships with a particular attribute; and / or the attribute model score {S k}1<=k<=j indicates that the performance of the ML technique has stabilized; or some other condition or criteria indicating that the selection model 106a has been effectively trained and can be used to select more candidate lists of compounds.

[0108] Figure 1d is another flow chart showing another example training process 130 for training a selection model to select a Figure 1c The selection model is generated by training the RL technique for predicting a candidate list of compounds that will enhance further training of the attribute model. The training process 130 may be implemented by the selection mechanism 106 and / or the training device 120. The selection model may initially be guided by using the training device 100 and / or the training process 110, as previously described in Figure 1a-1b In this case, the training process 130 can be used to further enhance the selection model and / or the attribute model is further enhanced at the same time. It should also be noted that the exemplary training process 130 can also be used to train the RL technique from scratch to generate a selection model for selecting a candidate list of compounds for validation and further enhance the training of the RL technique for generating the attribute model. The process 130 for training or retraining the RL technique to generate the selection model may include the following steps, wherein the selection model may be based on the result prediction list output from the attribute model Mj and / or the attribute model score S j To better predict the candidate list of compounds:

[0109] In step 132, the prediction result list {R l} jand the attribute model score S for the current iteration j j is received by the RL technique / selection model. In step 134, based on the attribute model score S j and / or the previous attribute model scores {S k} for 1 <= k < j, it is determined whether to retrain the selection model to select a set of compounds for the compound candidate list. For example, the attribute model scores {S k} for the previous iteration (e.g., k = j - 1) for 1 <= k < j can be compared with the attribute model score S j for the current iteration. If the performance of the attribute model has improved in terms of quality / accuracy, it is fed back as a reward to the RL technique, and the selection model can be retrained (e.g., "yes"), and proceed to step 135. Then, the attribute model can be retained / saved for another iteration of training the selection model. If in step 134, the comparison between the attribute model score S j and the previous attribute model scores {S k} for 1 <= k < j indicates that the quality / accuracy of the performance of the attribute model has not improved in the current iteration, it is fed back as a penalty to the RL technique, and the selection model can be retrained (e.g., "yes"), and then proceed to step 135. If it is determined based on the prediction result list {R l} j the attribute model score S j and / or the previous attribute model scores {S k} that retraining is not required, process 130 proceeds to step 136.

[0110] In step 135, depending on the type of retraining, the selection model associated with the RL technique can be updated / retrained based on the reward or penalty. If the selection model associated with the RL technique is to be updated / retrained based on the penalty, given that the performance of the attribute model has deteriorated, the selection model can indicate to the ML technique that it should revert to the previously retained / saved attribute model before the performance of the attribute model became worse. Once the selection model is retrained, process 130 can proceed to step 136.

[0111] In step 136, the selection model can be used to select from the attribute model M jA group of compounds are selected for the compound candidate list in the output prediction result list for verification of the compound candidate list. In step 138, the selected compound candidate list is sent by the selection model for verification and / or update of the attribute model. Verification of the selected compound candidate list can include the association of each compound in the compound candidate list with a specific attribute based on a labeled training data set, a laboratory experiment and / or a computer analysis verification. For example, a laboratory experiment can be performed on the compound candidate list to determine whether these compounds show or are associated with a specific attribute. The laboratory experiment can output a group of laboratory experiment verification results related to the selected compound candidate list and the attribute. The attribute model can be updated based on the laboratory experiment verification results and the ML technology for generating / updating the attribute model. Additionally or alternatively, a computer analysis (e.g., simulation) can be performed on the compound candidate list to determine whether these compounds show a specific attribute or are associated with a specific attribute. Computer analysis (e.g., simulation) can output a group of computer analysis verification results related to the selected compound candidate list and the specific attribute. The attribute model can be updated based on the computer analysis verification results and the ML technology for generating / updating the attribute model.

[0112] Computer analysis / simulation and / or laboratory experiments can be used to verify the relationship between the selected candidate compounds and specific attributes. Computer analysis / simulation can be used to verify whether each selected candidate compound has a specific attribute. Computer analysis (e.g., simulation) can output a set of computer analysis verification results related to the selected candidate list of compounds and specific attributes. Sometimes, as described herein, it can be determined to verify some or all of the selected candidate lists of compounds by laboratory experiments. The laboratory experiment can output a set of laboratory experiment verification results related to the selected candidate list of compounds and attributes. The attribute model can be updated based on ML technology, labeled training data sets, and computer analysis verification results or laboratory experiment verification results, wherein the laboratory experiment verification results verify the candidate list of compounds. That is, the verification results of the verified candidate list of compounds can be represented as a further labeled training data set associated with the candidate list of compounds, which can be used to further train the ML technology to generate / update the attribute model. Multiple compounds {C l}1<=l<=L is input to the updated attribute model, and a prediction result list {R l} j and attribute model score S j That is, it is possible to update the property model based on multiple compounds {C l}1<=l<=L to generate the attribute model score S j And further prediction result list {R l} j .

[0113] The selection model training process 130 may proceed to the next iteration (e.g., j=j+1) and step 132, where the selection model receives another result prediction list and / or attribute model score S from the updated attribute model. j .

[0114] In step 134, it may be determined that the selection model is fully trained and further training may not necessarily improve the selection of the candidate list of compounds. For example, if no improvement is seen in the predictive attribute model, the selection model may be considered to be trained and further training may be unnecessary. For example, one method of determining that the selection model is fully trained may include checking whether the selected candidate list of compounds sent to the laboratory and / or tested by computer simulation will make any subsequent predictive attribute model generated by retraining the ML technology based on laboratory or computer simulation results worse and / or the same. Comparing the previous attribute model score with the current retrained attribute model score can help determine whether the selection model can be considered to be fully trained. For example, when the retrained attribute model score is compared with the previous retrained attribute model score, indicating that the attribute model score is in a stable state, the selection model can be considered to be trained. When the currently updated / retained attribute model score is compared with the previously retained / saved attribute model score, indicating that the attribute model score is in a stable state, the selection model can be considered to be trained. In this case, the training process 130 may terminate and output a selection model for selecting a candidate list of compounds to augment the training of other property models configured to predict whether a compound is associated with a particular property.

[0115] As described above, the validated candidate list of compounds can then be used to enhance the training of the ML technique to generate a property model that predicts whether a compound exhibits or has a specific property. The specific property can be based on one of multiple properties associated with the compound. The selection process 100 is based on a reinforcement learning technique that iteratively trains a selection model to predict the best or most appropriate candidate list of compounds for validation by computer analysis and / or laboratory experiments / testing. The validation results can be used to further enhance the labeled training data set, which is used by the ML technique to generate and / or update the property model for a variety of compounds. Once trained, the selection model can be stored and used to predict a candidate list of compounds that need to be validated based on any specific property in a variety of properties.

[0116] Other modifications to process 130 may include that in response to determining to retrain the selected model in step 134, an instruction may be sent to the ML technique for generating an updated attribute model to revert the attribute model to the previous attribute model when the attribute model score does not meet the attribute model performance threshold compared to the corresponding previous attribute model score. Alternatively or additionally, in step 134, when the attribute model score indicates that the attribute model performance threshold is met or exceeded compared to the corresponding previous attribute model score, an instruction may be sent to the ML technique for generating an updated attribute model to retrain the attribute model Mj instead of replacing it with the previously trained attribute model.

[0117] Further modifications can be made so that the selection model can be trained by RL techniques to select not only a candidate list of compounds, but also a validation method using computer analysis / simulation and / or laboratory experiments. Considering the cost of conducting laboratory experiments, it may be preferable to set rules that penalize RL techniques when the selection model selects laboratory experiments as a validation method too early in the training process 130, or when the use of computer analysis / simulation still needs improvement.

[0118] Additional modifications may be made to process 130, such as, by way of example only and not limitation, in steps 134-135, when retraining of the selection model is to be performed, the selection model is retrained using the list of prediction results by displaying the differences between the selection model or the RL technique or laboratory and / or computer analysis / simulation validation results used to generate the selection model and the original attribute model prediction results. This may essentially allow the selection model or RL technique to learn which compounds, molecules, or chemicals the attribute model is most likely to make incorrect predictions about.

[0119] For example, in steps 134-135, when it is determined to retrain the selection model, the retraining may further include retraining the selection model based on the attribute model score and the prediction result list. This may include using the prediction result list to identify one or more differences between the verification results associated with the verification of the compound candidate list and the current or previous prediction result list output from the current or previous attribute model. Alternatively or additionally, retraining the selection model using the prediction result list may further include retraining the selection model based on the structure of the compound that the attribute model may have incorrectly predicted.

[0120] In other modifications, the process 130 for generating the selection model can be repeated until it is determined that the selection model has been effectively trained. For example, the process 130 can include further training or updating the selection model by iterating steps 132-138, which include: step 132, receiving a list of prediction results; steps 134-135, retraining or updating the selection model; step 136, selecting a candidate list of compounds; and step 138, sending the candidate list for verification, wherein the verification results are used to update the attribute model. In each iteration of these steps, when the process 130 is repeated for the next iteration, an updated attribute model from a previous or current iteration can be used. These steps 132-138 can be repeatedly performed until it is determined that the selection model has been effectively trained, or a stop criterion has been reached or satisfied. For example, by analyzing the convergence or divergence of the attribute model scores of the updated attribute models in each iteration, it can be determined that the selection model is effectively trained. When the attribute model score does not substantially change to one or more previous and recent attribute model scores when the selected candidate list of compounds is subjected to computer analysis and / or laboratory experiments, the selection model can be considered to be effectively trained. In another example, the stopping criteria may include that a maximum number of iterations has been reached. It will be appreciated by those skilled in the art that any other determinations / considerations / criteria / criteria may be used to determine the stopping criteria for selecting when the model is effectively trained and / or meets application requirements.

[0121] Figure 2 is a list of example prediction results {R l} j 200 tables for predicting multiple compounds {C l} is associated with a specific attribute, which can be used as an input to the selection model according to the present invention. l The associated attribute prediction value / score may include a prediction score P l The prediction result list {R l} j 200 includes multiple compounds {C l}202 and its corresponding prediction score when 1<=l<=L{P l}204 (e.g., attribute prediction value / score). Multiple compounds {C l} including compounds C1, C2, ..., C l , ..., C L-1 , C L The corresponding multiple prediction scores {P l}204 includes prediction scores P1, P2, ..., P l , ..., P L-1 , P L。Each prediction score Pl represents each of the compounds C l as to whether it has or is associated with a particular property. The verification step 106 can select a list of compound candidates from the list of prediction results {R l} j 200 based at least in part on the prediction scores.

[0122] As previously mentioned, the prediction score includes or represents data that represents or indicates a value of an ML model that predicts whether a compound has a particular property. The prediction score can be a value, by way of example only and not limitation, a probability value, a certainty value or score, a percentage score, or any other value that represents a prediction as to whether a compound has a particular property, or as to whether a compound exhibits a particular property, and / or a prediction as to the degree of association of a compound with a particular property; and / or any other value, score, or statistical data useful for evaluating or classifying whether a compound is associated with a particular property, etc.

[0123] For example, the prediction score P l as to whether compound C l is associated with a particular property can be represented as a certainty score value. Compounds known to have a particular property are assigned a value representing a "positive" certainty score (e.g., PCP). Compounds known not to have a particular property are assigned a value representing a "negative" certainty score (e.g., PCN). Other compounds are assigned a value representing an "uncertainty" score (Pl = Xl, where PCN < Xl < PCP). The "uncertainty" score can be a continuous real value that represents the level of uncertainty of the ML model as to whether the compound is associated with a particular property. The "uncertainty" score can have a continuous value between the value representing the positive certainty score and the value representing the negative certainty score (e.g., PCN < Pl < PCP). In this example, the certainty score is represented as a percentage certainty score, where the positive certainty score is 100%, the negative certainty score is 0%, and the uncertainty score is between the positive and negative certainty scores, i.e., between 0% and 100%.

[0124] In Figure 2 the list of prediction results {R l} j 200 sorts multiple compounds {C l}202 based on their prediction scores {P l}204. For example, if a compound has or exhibits a particular property, the prediction score can have a positive certainty level represented as a probability in the range of 1 or 100% (e.g., in the range of 0.85 - 1 or in the range of 85 - 100%). In Figure 2In it, the correct qualitative scores of C1 and C2 are expressed as PCP = 100%, which means that the ML model is 100% confident that these compounds C1 and C2 have specific properties. In addition, C L-1 and C L have negative certainty scores expressed as PCN = 0%, which means that the M1 model is 100% confident that these compounds C L-1 and C L do not have specific properties. There may be one or more compounds {C l}, the value of whose prediction score Pl = Xl is between PCN < Pl < PCP, where the ML model has a continuous confidence regarding whether these compounds are associated with specific properties. There are also compounds in the intermediate region between PCN and PCP (e.g., 45% < Pl < 55%), and the property model is most uncertain about whether these compounds are related to specific properties. It is precisely these compounds that may be selected into the compound candidate list and then undergo verification related to specific properties.

[0125] As an example, if a known compound has or indeed exhibits a specific property, the prediction score Pl of this compound can have a correct qualitative level, expressed as a probability within the range of 1 or 100% (e.g., within the range of 0.85 - 1 or 85 - 100%). If it is known that the compound does not have or does not exhibit a specific property, the prediction score Pl of this compound can have a negative certainty level, expressed as a probability within the range of 0 or 0% (e.g., within the range of 0 - 0.15 or 0 - 15%). Compounds with prediction scores between the correct qualitative level and the negative certainty level can be considered to have uncertain or ambiguous prediction scores. For example, those compounds with prediction scores within the range of 0.5 or 50% (e.g., between 0.45 and 0.55 or between 45 - 55%) can be considered the most uncertain or most ambiguous. That is to say, the property model cannot determine in one way or another whether these compounds have (exhibit or do not exhibit) specific properties. It is precisely these compounds that need to be verified for their relationship with specific properties and thus generate a further labeled training dataset for updating the property model as described herein.

[0126] A ruleset can be defined that selects the compounds considered to be the best from a list of prediction results. For example, by way of illustration only and not limitation, the most uncertain compounds are selected from the list of prediction results; the compounds that are structurally least similar compared to the compounds associated with the training dataset of labels used to train the property model, and many other combinations or rules. However, the ruleset for selecting the list of compound candidates may not produce the best selection of compounds that maximizes the quality of any updated property model. The training apparatuses and processes 100, 110, 120, and 130 do not define the ruleset for selecting the list of compound candidates, but rather define the generation of the selection model by training RL techniques to select a list of compound candidates that maximizes or enhances the quality and / or accuracy of the property model. By using RL techniques, the selection model can define a "ruleset" that learns all the necessary details required to select the best list of compound candidates at any given iteration of the training process, which should or will enhance the training of the property model.

[0127] Figure 3 FIG. is a schematic diagram showing an example selection mechanism 300 for selecting a list of compound candidates and a validation method according to the present invention. The selection mechanism 300 can be used to further modify or replace the selection mechanism 106 for selecting and validating a list of compound candidates for training ML techniques to generate or update a property model. The selection mechanism 300 includes a retraining mechanism 302, a selection model 304, a validation selector 306, a computer analysis validator 306a, and a laboratory validator 306b. The selection mechanism 300 receives at least the list of prediction results {R l} j 200, the property model score S j and / or the property model scores {S k}. The retraining mechanism 302 determines whether to retrain the selection model and the validation model at least partially based on the list of prediction results {R l} j 200, the property model score S j and / or the property model scores {S k}. The selection model 304 and the validation selector 306 can predict a list of compound candidates {C l} j and a validation method from the list of prediction results {R k} j 200. When the list of candidates is verified by the computer analysis validator 306a or the laboratory validator 306b based on a specific property, the list of compound candidates {C k} j and the validation method will enhance the update of the property model M j in the next iteration of the training process 110 or 130.

[0128] The verification selector 306 can be configured to indicate the selection of computer analysis / simulation via a selector VT or some other technique / method, so that the candidate list of compounds {Ck}j is indicated / requested to be processed by a computer analysis verifier 306a, which is used to verify the candidate list of compounds. The computer analysis verifier 306a can be connected to one or more computer analysis / simulation systems [e.g., molecular dynamics (MD) (RTM) molecular simulators], which can atomically simulate whether a compound has or exhibits a specific property. For example, an MD simulator uses atomic and / or physical simulations of molecules to simulate the properties of a compound / molecule. The types of properties of compounds that can be simulated by MD include, by way of example only but not limited to, docking simulations, including protein docking with compounds, and / or any other properties or compounds that can be simulated to determine whether a compound has a specific property.

[0129] The computer analysis / simulator validator 306a validates the candidate list by sending the candidate list to the computer analysis / simulation system, which verifies the candidate list based on the specific attributes and compound candidate list {C k} j The computer analysis / simulation analysis can be performed by the computer analysis / simulation verifier 306a. The computer analysis / simulation result can be received from the computer analysis / simulation system. The computer analysis / simulation result can be used to estimate the association between each compound on the candidate compound list and a specific attribute. k} j The associated computer analysis / simulation results can be obtained by using a labeled training data set {T k} j C The training dataset can be used to generate further training datasets {T k} j , as described in this paper, is used to generate / update the attribute model Mj through ML techniques. The selector VT can be used to select a labeled training dataset {T k} j C As a further training dataset {T k} j , used to train ML techniques to generate / update attribute models M j .

[0130] Verification selector 306 may be configured to indicate the selection of laboratory experiments by selector VT or some other technique / method so that the candidate list of compounds {C k} jThe laboratory validator 306b may be connected to one or more computer systems associated with one or more laboratories that may receive the compound candidate list and perform laboratory experiments related to whether each compound in the candidate list has or exhibits a specific property. k} j The associated experimental results can be obtained with a labeled training dataset {T k} j L Output in the form of .

[0131] Optionally, the laboratory validator 306b can notify the operator of the candidate list of compounds and specific attributes for laboratory experiments. The operator can send the candidate list of compounds and ask the laboratory to perform experiments to determine whether each of the candidate list of compounds has or exhibits a specific attribute. After the experiment is completed, the experimental results and / or further training data associated with the candidate list of compounds and whether each compound has or exhibits a specific attribute can be sent to the laboratory validator 306b.

[0132] The laboratory validator 306b may be configured to output a labeled training data set {T k} j L As described in this paper, the labeled training dataset {T k} j L Can be used as further training data {T k} j , used to generate / update the attribute model M through ML techniques j The selector VT can be used to select the labeled training dataset {T k} j L As a further training dataset {T k} j , used to train ML techniques to generate / update attribute models M j .

[0133] Although the selector VT is shown as a switching circuit that switches between the computer analysis / simulator verifier 306a and the laboratory verifier 306b, this is merely an example and the present invention is not limited thereto. It should be understood that a person skilled in the art may use any other method, technique, device or hardware / software to select and / or direct / request the computer analysis / simulator verifier 306a and / or the laboratory verifier 306b to process the candidate list of compounds based on specific attributes.

[0134] The retraining mechanism 302 can implement the retraining step of the training process 110 or 130, which can be further modified to train the RL technology to generate a selection and / or validation model that selects a candidate list of compounds and a validation method for validating the selected candidate list of compounds. The validation selector 306 can implement the validation model generated by the RL technology, but when the RL technology is retrained in the retraining mechanism 302, several rules for rewarding or penalizing the RL technology can be included. For example, given that laboratory experiments are costly and time-consuming, when generating a selection model for selecting a candidate list and a validation method, the validation selector 306 can prohibit the selection of performing laboratory experiments too early in the training process.

[0135] As an example, when the verification selector 306 selects to perform a laboratory experiment and the number of iterations for retraining the selection / verification model is below a predetermined threshold, the verification penalty 308a (e.g., PNj+1) can be fed back to the retraining mechanism 302 in the next iteration (e.g., j=j+1) for retraining the selection / verification method based on the penalty. The verification selector 306 can temporarily prohibit the selection of using the laboratory verifier 306b for verification, and instead select the verification method of performing computer analysis using the computer analysis verifier 306a.

[0136] In another example, when the verification selector 306 selects to perform laboratory experiments and determines that the verification method of performing computer analysis will further improve the attribute model score, the verification penalty 308a (e.g., PNj+1) can be fed back to the retraining mechanism 302 in the next iteration (e.g., j=j+1) for retraining the selection / verification method based on the penalty. This is to encourage the selection / verification model to use computer analysis to verify the candidate list as much as possible, minimizing the cost and expense of laboratory experiments. The verification selector 306 can temporarily prohibit the selection of using the laboratory verifier 306b for verification, and instead select the verification method of performing computer analysis using the computer analysis verifier 306a.

[0137] In another example, when the validation selector 306 selects to perform laboratory experiments and the selected candidate list of compounds is already significantly different from the previously selected candidate list of compounds, the validation penalty 308a (e.g., PNj+1) can be fed back to the retraining mechanism 302 in the next iteration (e.g., j=j+1) for retraining the selection / validation method based on the penalty. This is to encourage the selection / validation model to use computer analysis to validate the candidate list when the candidate list changes between iterations, because this indicates that computer analysis is still useful for validating the candidate list and further minimizes the cost and expense of laboratory experiments. The validation selector 306 can temporarily prohibit the selection of using the laboratory validator 306b for validation, and instead select a validation method that performs computer analysis using the computer analysis validator 306a.

[0138] In another example, the validation selector 306 may select to perform computer analysis while determining that the computer analysis will improve the attribute model score of the attribute model based on the previous attribute model score, the previous attribute model score being calculated based on the corresponding predicted result list generated after each compound candidate list has been validated. The validation reward 308b (e.g., RWj+1) can be fed back to the retraining mechanism 302 in the next iteration (e.g., j=j+1) for retraining the selection / validation method based on the reward. The validation selector 306 continues to allow the validation method to perform computer analysis using the computer analysis validator 306a.

[0139] In another example, the validation selector 306 may select to perform laboratory analysis while determining that the computer analysis will not improve the attribute model score of the attribute model based on the previous attribute model score, the previous attribute model score being calculated based on the corresponding predicted result list generated after each compound candidate list has been validated; and determining that the compound candidate list is small enough for laboratory analysis. The validation reward 308b (e.g., RWj+1) can then be fed back to the retraining mechanism 302 in the next iteration (e.g., j=j+1) for retraining the selection / validation method based on the reward. The validation selector 306 continues to allow the validation method to perform laboratory analysis on the selected compound candidate list using the laboratory analysis validator 306b.

[0140] Although several examples of penalties or rewards are provided as examples but not limitation to represent the feedback signals of the verification penalty 308a (e.g., PNj+1) or the verification reward 308b (e.g., RWj+1) for retraining the selection model / validation model in the next iteration, it should be understood by those skilled in the art that any other method / mechanism for retraining the selection / validation model may be used, and the verification selector 306 may provide the penalty or reward as a continuous or discrete value to the retraining mechanism at any time. For example, the retraining mechanism may perform a micro-retraining loop iteration within each main iteration j. Each micro-retraining loop may include, by way of example but not limitation, the retraining mechanism 302, the candidate list selector 304, and the verification selector 306, wherein a plurality of micro-iterations i may be performed for each iteration j, wherein the verification selector 306 feeds back the verification penalty 308a (e.g., PNi) or the verification reward 308b (e.g., RWi) to the retraining mechanism 302 for retraining the selection model / validation model to select an appropriate verification method before executing the next iteration j+1.

[0141] Additionally or alternatively, the selection of verification methods (e.g., computer analysis verifier 306a and / or laboratory analysis verifier 306b) can be performed heuristically, for example, the verification selector 306 can choose to use the computer analysis / simulation of the computer analysis verifier 306a to verify the candidate list of compounds. Once the attribute model stops improving based on the computer analysis / simulation verification, the verifier selector 306 can choose to verify the appropriately selected candidate list of compounds using the laboratory experiment of the laboratory analysis verifier 306a. The experimental results can be used to update the attribute model, wherein the computer analysis / simulation using the computer analysis verifier 306a can be selected again until the attribute model stops improving. Regardless of selecting computer simulation and / or laboratory experiment, this can be repeated until the performance of the attribute model stabilizes or stops improving. Alternatively or additionally, this can be repeated a predetermined number of times and / or until a predetermined number of laboratory experiments have been performed, and / or a predetermined number of computer simulations have been performed, and / or a combination of the two.

[0142] This heuristic technique can also be combined with the training method used to train the selection model to select an appropriate verification method to verify the candidate list of compounds. Heuristic techniques can be used to reward and / or punish the selection model for selecting an inappropriate verification method, and can therefore be used to intervene and ensure that an appropriate verification method is selected. For example, when computer analysis / simulation verification can still produce an improved attribute model, and / or when the selection model is still being trained to select an appropriate candidate list of compounds for verification, this can be used to prevent the selection model from prematurely selecting the laboratory analysis verifier 306b for laboratory experiments.

[0143] Figure 44 is a schematic diagram of a computing system 400 including a computing device or apparatus 402 according to the present invention. The computing device or apparatus 402 may include a processor unit 404, a memory unit 406, and a communication interface 408. The processor unit 404 is connected to the memory unit 406 and the communication interface 408. The memory unit 406 may include an operating system (OS) and a data storage (DS), which may include other applications and / or software, such as, by way of example only and not limitation, computer-implemented methods, processes, and / or methods for implementing reference Figures 1a to 3 The processor unit 404 and the memory unit 406 may be configured to implement the methods and / or processes described in the process 110, 130 and / or as described in the reference Figures 1a to 3 The processor unit 404 may include one or more processors, controllers, or any suitable type of hardware for implementing computer-executable instructions to control the device 402 according to the present invention. The computing device 402 may be connected to a network 412 via a communication interface 408 for communicating and / or operating with other computing devices / systems (not shown) to implement the present invention accordingly.

[0144] The computing system 400 may be a server system that may include a single server or a server network configured to implement the present invention as described herein. In some examples, the functionality of the server may be provided by a server network distributed over a geographic region, such as a globally distributed server network, and a user may be connected to an appropriate one of the server networks based on the user's location.

[0145] Further modifications or examples may include a computer-implemented method for updating an attribute model for predicting whether a compound is associated with a particular attribute, wherein the method comprises the following steps: generating a result list of compounds using the attribute model of a plurality of compounds; selecting a candidate list of compounds from the plurality of compounds using the selection model according to the process 110, 130 and / or the apparatus / system 100, 120, 300 and / or modifications thereof and / or as described herein; receiving validation results of the candidate list of compounds (e.g., from a computer analysis, a laboratory experiment, or a labeled training data set); and updating the attribute model based on the validation results. These steps may be repeated until it is determined that the attribute model is effectively trained.

[0146] The method may include validating an association of each of the candidate list of compounds with a particular attribute, wherein the validation output includes validation results representing data of further labeled training data corresponding to the validated attribute association of each compound in the candidate list of compounds. The attribute model may also be updated based on training the ML technique based on labeled training data sets corresponding to a plurality of compounds and their associations with a particular attribute.

[0147] An apparatus or computing device 402 comprises a processor 404, a memory unit 406 and a communication interface 408, wherein the processor 404 is connected to the memory unit 406 and the communication interface 408, wherein the processor 404, the communication interface 408 and / or the memory unit 406 are configured to implement a computer-implemented method for updating a property model.

[0148] Other modifications or examples may include a system for generating a selection model based on an ML technique (e.g., an RL technique or any other ML technique), the selection model being configured to select a candidate list of compounds for verification related to a specific attribute. The system may include: a selection module or device configured according to any of the processes 110, 130 and / or the apparatus / system 100, 120, 300 and / or modifications thereof and / or as described herein, for selecting a candidate list of compounds; and a property update module or device configured according to a method for updating a property model and / or the apparatus, modifications thereof and / or computer implementations as described herein, wherein the property update module or device is coupled to the selection module, and the property update module or device is configured to update the property model based on the selected candidate list of compounds.

[0149] In a further modification, a system may be provided, the system comprising: an attribute model generation module configured to generate an attribute model based on a labeled training data set, the labeled training data set including data representing compounds associated with a specific attribute; an ML verification module configured to generate a prediction result list output from the generated attribute model for predicting whether a plurality of compounds are associated with a specific attribute and an attribute model score associated with the prediction; and a selection module, which is configured according to any one of the processes 110, 130 and / or the apparatus / system 100, 120, 300 and / or modifications thereof and / or as described herein, for selecting a candidate list of compounds for verification from the prediction result list, wherein the verification compound candidate list outputs a verification result for updating the attribute model. The attribute model generation module may also be configured to receive a further labeled training data set based on the verification result of the candidate list of compounds, and update the attribute model by generating an attribute model based on the labeled training data set and the further labeled training data set. The system may include the processes 110 , 130 and / or the apparatus / systems 100 , 120 , 300 and / or modifications thereof and / or one or more further modifications, features, steps and / or features as described herein.

[0150] In addition, refer to Figure 1a-4 One or more of the devices and / or systems 100, 120, 300, processes 110 and / or 130 and / or methods / processes described in can be implemented in hardware and / or software. Figure 1a-4One or more of the methods and / or processes described in for training and / or implementing a selection model and / or for using a selection model may be implemented in hardware and / or software, for example, but not limited to, as a computer-implemented method executed by one or more processors / processor units or as an application requirement. Such an apparatus, system, process and / or method may be used to generate an ML model that includes data representing a selection model, the selection model being described in reference to Figure 1a-4 , modifications thereof, and / or the apparatus and / or system 100, 120, 300, process 110 and / or 130 and / or apparatus, system and / or method / process as described herein, etc. Thus, the ML selection model can be obtained from the apparatus, system and / or computer-implemented process, method as described herein. In addition, the ML validation model can also be obtained from the apparatus, system and / or computer-implemented process, method as described herein. Figure 1a-4 One or more of the devices and / or systems 100, 120, 300, processes 110 and / or 130 and / or devices, systems and / or methods / processes described in the foregoing are obtained, some of which may be implemented in hardware and / or software, such as, by way of example only but not limitation, computer-implemented methods that may be executed on a processor or processor unit, or as required by the application, such as reference Figure 1a-4 One or more of, modifications thereof, and / or as described herein, and the like.

[0151] For the sake of clarity, the above description discusses the embodiments of the present invention based on a single user. It should be understood that in practice, the system can be shared by multiple users, and may be shared by a large number of users at the same time.

[0152] The above embodiments are fully automatic. In some examples, a user or operator of the system may manually instruct some steps of the method to be performed.

[0153] In the described embodiments of the present invention, the system may be implemented as any form of computing and / or electronic device. Such a device may include one or more processors, which may be microprocessors, controllers, or any other suitable type of processor, for processing computer executable instructions to control the operation of the device so as to collect and record routing information. In some examples, such as where a system-on-chip architecture is used, the processor may include one or more fixed function blocks (also referred to as accelerators) that implement a portion of the method in hardware (rather than software or firmware). Platform software including an operating system or any other suitable platform software may be provided at a computing-based device to enable application software to be executed on the device.

[0154] The various functions described herein can be implemented with hardware, software or any combination thereof. If implemented with software, these functions can be stored on or transmitted by a computer-readable medium as one or more instructions or codes. Computer-readable media may include, for example, computer-readable storage media. Computer-readable storage media may include volatile or non-volatile, removable or non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer-readable storage media may be any available storage medium accessible by a computer. As an example and not limitation, such computer-readable storage media may include RAM, ROM, EEPROM, flash memory or other storage devices, CD-ROM or other optical disk storage devices, magnetic disk storage devices or other magnetic storage devices, or any other medium that can be used to carry or store desired program codes in the form of instructions or data structures and can be accessed by a computer. The disks used herein include compact discs (CDs), laser discs, optical discs, digital versatile discs (DVDs), floppy disks and blue-ray discs (BDs). In addition, propagated signals are not included in the scope of computer-readable storage media. Computer-readable media also include communication media, including any media that facilitates the transfer of computer programs from one place to another. For example, a connection may be a communication medium. For example, if the software is transmitted from a website, server or other remote source using coaxial cable, fiber optic cable, twisted pair, DSL or wireless technology (such as infrared, radio and microwave), it belongs to communication media. Combinations of the above should also be included in the scope of computer-readable media.

[0155] Alternatively or additionally, the functions described herein may be at least partially performed by one or more hardware logic components. For example, but not limited to, the hardware logic components that may be used may include field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chip (SOCs), complex programmable logic devices (CPLDs), etc.

[0156] Although illustrated as a single system, it should be understood that the computing device may be a distributed system. Thus, for example, several devices may communicate via a network connection and may jointly perform the tasks described as being performed by the computing device.

[0157] Although illustrated as a local device, it will be appreciated that the computing device may be located remotely, accessed via a network or other communications link (eg, using a communications interface).

[0158] The term "computer" is used herein to refer to any device having processing power so as to be able to execute instructions. Those skilled in the art will appreciate that such processing power is incorporated into many different devices, and thus the term "computer" includes PCs, servers, mobile phones, personal digital assistants, and many other devices.

[0159] Those skilled in the art will appreciate that storage devices for storing program instructions can be distributed over a network. For example, a remote computer can store examples of processes described as software. A local or terminal computer can access the remote computer and download some or all of the software to run the program. Alternatively, the local computer can download software fragments as needed, or execute some software instructions at a local terminal, or execute some software instructions at a remote computer (or computer network). Those skilled in the art will also appreciate that all or part of the software instructions can be executed by a dedicated circuit, such as a DSP, a programmable logic array, etc., by utilizing conventional techniques known to those skilled in the art.

[0160] It should be understood that the above benefits and advantages may relate to one embodiment or may relate to several embodiments. The embodiments are not limited to embodiments that solve any or all of the problems described, or embodiments that have any or all of the benefits and advantages described. Variants should be considered to be included within the scope of the present invention.

[0161] Any reference to "an" object refers to one or more of said object. The term "comprising" is used herein to indicate including the identified method steps or elements, but these steps or elements do not include an exclusive list and the method or apparatus may contain additional steps or elements.

[0162] As used herein, the terms "component" and "system" are intended to include a computer-readable data storage configured with computer-executable instructions that, when executed by a processor, cause certain functions to be performed. Computer-executable instructions may include routines, functions, etc. It should also be understood that a component or system may be located on a single device or distributed across several devices.

[0163] Additionally, the term "exemplary," as used herein, is intended to mean "serving as an illustration or example of something."

[0164] Furthermore, the term "comprising" used in the detailed description or claims should be interpreted similarly to the term "including" when used as a transitional word in a claim.

[0165] The accompanying drawings illustrate exemplary methods. Although these methods are shown and described as a series of actions performed in a particular order, it should be understood and appreciated that these methods are not limited by the order. For example, some actions may occur in an order different from that described herein. In addition, an action may occur simultaneously with another action. In addition, in some cases, all actions may not be required to implement the methods described herein.

[0166] In addition, the actions described herein may include computer-executable instructions that can be implemented by one or more processors and / or stored on a computer-readable medium. The computer-executable instructions may include routines, subroutines, programs, execution threads, etc. In addition, the results of the actions of these methods may be stored in a computer-readable medium, displayed on a display device, etc.

[0167] The order of steps of the methods described herein is exemplary, and the steps may be performed in any suitable order, or simultaneously where appropriate. In addition, steps may be added or substituted in any method, or individual steps may be deleted from any method, without departing from the scope of the subject matter described herein. Aspects of any of the above examples may be combined with aspects of any other example described to form further examples without losing the effect sought.

[0168] It should be understood that the above description of the preferred embodiment is given as an example only, and various modifications may be made by those skilled in the art. The above description includes examples of one or more embodiments. Of course, in order to describe the above aspects, it is impossible to describe every conceivable modification and change of the above-mentioned device or method, but it will be appreciated by those of ordinary skill in the art that many further modifications and permutations of various aspects are possible. Therefore, the described aspects are intended to include all such changes, modifications and variations that fall within the scope of the appended claims.

Claims

1. A computer-implemented method for generating a selection model to select a candidate list of compounds for specific attribute-related validation, the method comprising: receiving a prediction result list output from a property model for predicting whether each of a plurality of compounds has the specific property, and calculating a property model score based on the prediction result list, wherein the property model score estimates the quality of the property model; generating the selection model by training a machine learning (ML) technique based on the attribute model scores; training the selection model based on the attribute model scores; selecting a candidate list of compounds from the plurality of compounds for validation using the trained selection model, wherein the compounds are selected from the list of predicted results; sending the selected candidate list of compounds to verify the specified properties; verifying whether each compound in the selected candidate list of compounds has the specific attribute, wherein the attribute model is updated based on the verification; as well as The receiving and training of the selection model are repeated at least until it is determined that the selection model has been effectively trained, wherein the selection model is determined to be effectively trained when the attribute model score has not changed to one or more previous and most recent attribute model scores. 2 . The computer-implemented method of claim 1 , wherein training the selection model further comprises determining whether to train the selection model to select a candidate list of compounds based on the attribute model score and a previous attribute model score.

3. A computer-implemented method according to claim 1, wherein the ML technique generates the property model using a labeled training data set, the labeled training data set is associated with a subset of the plurality of compounds related to the specific property, and wherein the validation candidate list of the compounds is incorporated into the labeled training data set to update the property model by the ML technique. 4 . The computer-implemented method of claim 1 , wherein validating the selected candidate list of compounds further comprises validating that each compound from the candidate list of compounds has the specific property based on laboratory experiments or computer analysis.

5. The computer-implemented method of claim 4, wherein The laboratory experiment outputs a set of laboratory experiment validation results related to the selected compound candidate list and the attribute, wherein the attribute model is updated based on the laboratory experiment validation results and the ML technique used to generate the attribute model; or The computer analysis outputs a set of computer analysis validation results associated with the selected candidate list of compounds and the specific attribute, wherein the attribute model is updated based on the computer analysis validation results and the ML technique used to generate the attribute model.

6. The computer-implemented method of claim 1 , wherein the selection model is further trained to select a validation method for validating the selected candidate list of compounds, wherein selecting the validation method for validating the selected candidate list of compounds further comprises: selecting whether to conduct laboratory experiments or computer analysis based on the specific attributes and the selected candidate list of compounds; performing a laboratory experiment in response to the selection, wherein the laboratory experiment outputs a laboratory experiment verification result for estimating whether each compound on the selected candidate list of compounds has the specific property, wherein the laboratory experiment verification result is used to update the property model; In response to the determination, a computer analysis is performed, wherein the computer analysis outputs a computer analysis verification result for estimating whether each compound on the selected candidate list of compounds has the specific property, wherein the computer analysis verification result is used to update the property model.

7. The computer-implemented method of claim 5, further comprising: receiving a prediction result list output from the property model for predicting whether a plurality of compounds have the specific property and a property model score; determining whether to train the selection model to select a candidate list of compounds and a validation method based on the attribute model score and a previous attribute model score; as well as The selection model is trained based on the attribute model scores and / or the prediction result list.

8. The computer-implemented method of claim 6, wherein when a validation method of performing a laboratory experiment is selected and the number of iterations for training the selected model is below a predetermined threshold, the method further comprises: penalizing the selection model during training; as well as Select the validation method to perform the in silico analysis.

9. The computer-implemented method of claim 6, wherein when a validation method of performing laboratory experiments is selected and it is determined that a validation method of performing computer analysis will further improve the attribute model score, the method further comprises: penalizing the selection model during training; as well as Select the validation method to perform the in silico analysis.

10. The computer-implemented method of claim 6, wherein when a validation method of performing laboratory experiments is selected and the selected candidate list of compounds has substantially changed from a previously selected candidate list of compounds, the method further comprises: penalizing the selection model during training; as well as Select the validation method to perform the in silico analysis.

11. The computer-implemented method of claim 6, wherein when a validation method is selected that performs computer analysis and it is determined that the computer analysis will improve the attribute model score of the attribute model based on a previous attribute model score calculated from a corresponding list of predicted results generated after each compound candidate list has been validated, the method further comprises: rewarding the selection model during training; as well as Select the validation method to perform the in silico analysis.

12. The computer-implemented method of claim 1, wherein the prediction results list includes a predicted property score indicating whether each of the compounds has the particular property.

13. A computer-implemented method according to claim 12, wherein the predicted attribute score includes a certainty score, wherein compounds known to have the specific attribute are given a positive certainty score, compounds known not to have the specific attribute are given a negative certainty score, and other compounds are given an uncertainty score between the positive certainty score and the negative certainty score.

14. The computer-implemented method of claim 13, wherein the certainty score is a percentage certainty score, wherein the positive certainty score is 100%, the negative certainty score is 0%, and the uncertainty score is between the positive and negative certainty scores.

15. The computer-implemented method of claim 1 , wherein training the selection model further comprises: instructing an ML technique associated with the attribute model to revert the attribute model to a previous attribute model when the attribute model score does not meet an attribute model performance threshold as compared to the corresponding previous attribute model score; instructing the ML technique associated with the attribute model to retain the updated attribute model over the previous attribute model when the attribute model score indicates that the attribute model performance threshold is met or exceeded as compared to the corresponding previous attribute model score; training the selection model to select a set of compounds from the corresponding list of prediction results based on the attribute model scores; as well as Repeat the steps of claim 1 until it is determined that the selected model is effectively trained.

16. The computer-implemented method of claim 15, wherein determining that the selection model is effectively trained further comprises: comparing the retained attribute model scores to previously retained attribute model scores; as well as A plateau in the attribute model scores is based on determining that the selection model has been effectively trained.

17. The computer-implemented method of any one of claims 1 to 16, wherein the ML technique used to generate the selection model comprises at least one ML technique or a combination of ML techniques from: Recurrent Neural Networks; Convolutional Neural Networks; Reinforcement learning algorithms; and / or Any neural network architecture.

18. The computer-implemented method of any one of claims 1 to 16, wherein the specific attribute comprises a property or characteristic indicating: The compound docks with another compound to form a stable complex; Ligand docking target protein, wherein the compound is the ligand; The compound docks or binds to one or more target proteins; Compounds have a specific solubility or solubility range; The compound has a specific toxicity; any other properties or characteristics associated with chemical compounds that can be simulated based on computer simulations and the physical motion of atoms and molecules; any other properties or characteristics associated with the compound that can be determined from the expert knowledge base; as well as Any other property or characteristic associated with a compound that can be determined experimentally.

19. The computer-implemented method of any one of claims 1 to 16, wherein the ML technique used to generate the selection model is different from the ML technique used to generate or update the attribute model.

20. The computer-implemented method of any one of claims 1 to 16, wherein the step of training the selection model further comprises training the selection model based on the attribute model scores and the list of predicted results.

21. A computer-implemented method according to any one of claims 1 to 16, wherein training the selection model based on the prediction result list further comprises training the selection model based on one or more differences between validation results associated with validation of the compound candidate list and the current or previous prediction result list output from the current or previous attribute model.

22. The computer-implemented method of claim 21, wherein using the list of prediction results to train the selection model comprises training the selection model based on compound structures that are likely to be incorrectly predicted by the attribute model.

23. A device comprising a processor, a memory unit and a communication interface, wherein the processor is connected to the memory unit and the communication interface, wherein the processor, the communication interface and / or the memory are configured to implement the computer-implemented method according to any one of claims 1 to 22.

24. A computer-readable medium comprising data or instruction code representing a machine learning (ML) selection model, wherein the machine learning selection model is generated based on training an ML technique according to a computer-implemented method according to any one of claims 1 to 22, and when executed on a processor, the computer-readable medium causes the processor to implement the ML selection model.

25. A device for generating a selection model based on machine learning (ML) technology, wherein the selection model is configured to select a candidate list of compounds for specific attribute-related verification, wherein the device is configured to: receiving a prediction result list output from a property model for predicting whether each of a plurality of compounds has a specific property, and calculating a property model score based on the prediction result list, wherein the property model score estimates the quality of the property model; generating the selection model by training an ML technique based on the attribute model scores; training the selection model based on the attribute model scores; selecting a candidate list of compounds from the plurality of compounds for validation using the trained selection model, wherein the compounds are selected from the list of predicted results; sending the selected candidate list of compounds to verify the specified properties; verifying whether each compound in the selected candidate list of compounds has the specific attribute, wherein the attribute model is updated based on the verification; as well as The receiving and training of the selection model are repeated at least until it is determined that the selection model has been effectively trained, wherein the selection model is determined to be effectively trained when the attribute model score has not changed to one or more previous and most recent attribute model scores.

26. A computer-implemented method for updating a property model, the property model being used to predict whether a compound is associated with a particular property, the method comprising: generating a result list of compounds using the property model of the plurality of compounds; selecting a candidate list of compounds from the plurality of compounds using a selection model according to any one of claims 1 to 22; receiving a verification result of the candidate compound list; as well as The attribute model is updated based on the verification result.

27. The computer-implemented method of claim 26, further comprising repeating the steps of generating, selecting, and receiving until it is determined that the attribute model is effectively trained.

28. The computer-implemented method of claim 26 or 27, further comprising: Verifying whether each compound in the candidate compound list has the specific attribute, wherein the verification outputs a verification result, the verification result including data representing further labeled training data, the further labeled training data corresponding to the verified attribute of each compound in the candidate compound list.

29. The computer-implemented method of any one of claims 26 to 27, wherein the property model is updated based on training a machine learning technique based on a training data set corresponding to a plurality of compounds and labels indicating whether the plurality of compounds have a particular property.

30. A device comprising a processor, a memory unit and a communication interface, wherein the processor is connected to the memory unit and the communication interface, wherein the processor, the communication interface and / or the memory unit are configured to implement the computer-implemented method according to any one of claims 26 to 29.

31. A system for generating a selection model based on machine learning (ML) technology, wherein the selection model is configured to select a candidate list of compounds for specific attribute-related verification, the system comprising: A selection module or a device according to claim 23 or 25, for selecting a candidate list of compounds; as well as A property updating module or device according to claim 30, the property updating module or device is coupled to the selection module, wherein the property updating module or device is configured to update the property model based on the selected compound candidate list.

Images