Molecular optimization methods, related apparatuses, and media

Abstract

The present disclosure provides a molecular optimization method, related device and medium. The method comprises: determining a first biomolecule and a second biomolecule; based on the first biomolecule, performing optimization prediction on the second biomolecule by a molecular optimization model to determine candidate biomolecules predicted in each round, the candidate biomolecules in each round being predicted based on the candidate biomolecules in the previous round relative to the round; performing reward prediction on the candidate biomolecules based on the prediction model to obtain reward prediction results of the candidate biomolecules; and based on the reward prediction results, determining a target biomolecule from the plurality of candidate biomolecules, the binding affinity of the target biomolecule and the first biomolecule being greater than the binding affinity of the second biomolecule and the first biomolecule. The present disclosure can improve the accuracy of molecular optimization and generate biomolecules with better binding affinity to the first biomolecule. The present disclosure can be applied to various scenarios such as big data, artificial intelligence, cloud technology, etc.

Description

Technical Field

[0001] This disclosure relates to the field of big data technology, and in particular to a molecular optimization method, related apparatus and medium. Background Technology

[0002] Currently, generating peptides that can bind to specific target proteins and generating antibodies that match a particular antigen are particularly important in biomedical scenarios such as drug development.

[0003] Although generative and optimization schemes for biomolecules such as peptides and antibodies have emerged in related technologies, these technologies are often based on generative autoencoders to generate specific peptide types or antibody types. They cannot generate peptides that bind to arbitrary specific target proteins, nor can they meet the requirement of generating corresponding antibodies against arbitrary antigens. Furthermore, the optimization of biomolecules such as peptides and antibodies in these technologies relies heavily on training optimization models with large-scale data, which can lead to poor optimization results and affect the binding performance of the optimized peptides and antibodies. Therefore, how to generate biomolecules with good binding affinity to specific target proteins and antigens remains a pressing problem to be solved. Summary of the Invention

[0004] This disclosure provides a molecular optimization method, related apparatus, and medium that can improve the accuracy of molecular optimization, thereby generating biomolecules with better binding affinity to the first biomolecule.

[0005] According to one aspect of this disclosure, a molecular optimization method is provided, the method comprising:

[0006] Identify the first biomolecule and the second biomolecule to be optimized;

[0007] Based on the first biomolecule, the second biomolecule is optimized and predicted using a preset molecular optimization model to determine candidate biomolecules predicted in each round, wherein the candidate biomolecules in each round are predicted based on the candidate biomolecules in the previous round relative to the round.

[0008] Reward predictions are performed on multiple candidate biomolecules based on a pre-trained prediction model, and reward prediction results for each of the multiple candidate biomolecules are obtained.

[0009] Based on the reward prediction results, a target biomolecule is identified among the plurality of candidate biomolecules, wherein the binding affinity between the target biomolecule and the first biomolecule is greater than the binding affinity between the second biomolecule and the first biomolecule.

[0010] According to one aspect of this disclosure, a molecular optimization apparatus is provided, the apparatus comprising:

[0011] The first determining unit is used to determine the first biomolecule and the second biomolecule to be optimized;

[0012] An optimization unit is configured to optimize and predict the second biomolecule based on the first biomolecule using a preset molecular optimization model, and determine candidate biomolecules predicted in each round, wherein the candidate biomolecules in each round are predicted based on the candidate biomolecules in the previous round relative to the round.

[0013] The prediction unit is used to perform reward prediction on multiple candidate biomolecules based on a pre-trained prediction model, and obtain the reward prediction results for each of the multiple candidate biomolecules.

[0014] The second determining unit is used to determine the target biomolecule among the plurality of candidate biomolecules based on the reward prediction result, wherein the binding affinity between the target biomolecule and the first biomolecule is greater than the binding affinity between the second biomolecule and the first biomolecule.

[0015] Optionally, the optimization unit includes:

[0016] The initialization module is used to initialize the current round number to 1 and the biomolecule to be examined to the second biomolecule.

[0017] The first determining module is used, in the current round, to optimize and predict the biomolecule to be examined based on the first biomolecule using the molecular optimization model, and to determine multiple molecular prediction results and the action gain data of each of the multiple molecular prediction results.

[0018] The second determining module is used to determine the candidate biomolecule from the plurality of molecular prediction results based on the action gain data;

[0019] The loop module is used to increment the round number by 1, update the biomolecule to be examined to the candidate biomolecule, and return to the step of optimizing and predicting the biomolecule to be examined based on the first biomolecule in the current round using the molecular optimization model, determining multiple molecular prediction results and the action gain data of each of the multiple molecular prediction results, until the round number is the target number, thus obtaining multiple candidate biomolecules.

[0020] Optionally, the first determining module includes:

[0021] The extraction submodule is used to extract the molecular feature information of the first biomolecule in the current round;

[0022] The prediction submodule is used to optimize and predict the biomolecule under investigation based on the molecular feature information and the molecular optimization model to obtain the predicted biomolecule, and to determine the action gain data of the predicted biomolecule based on the difference between the predicted biomolecule and the second biomolecule.

[0023] The loop submodule is used to update the biomolecule to be examined with the predicted biomolecule and return the steps of optimizing the prediction of the biomolecule to be examined based on the molecular feature information and through the molecular optimization model to obtain the predicted biomolecule, until it is determined that the predicted biomolecule and the action gain data meet the first condition, the predicted biomolecule is determined as the molecular prediction result, and the action gain data of the predicted biomolecule is determined as the action gain data of the molecular prediction result.

[0024] Optionally, the prediction submodule is used for:

[0025] Based on the difference between the predicted biomolecule and the second biomolecule, the action space matrix for transforming from the second biomolecule to the predicted biomolecule is determined.

[0026] Based on the action space matrix and the molecular feature information, the action gain data is obtained by performing gain prediction through the molecular optimization model.

[0027] Optionally, the loop submodule is used for:

[0028] Determine the action gain data of the biomolecule under investigation corresponding to the predicted biomolecule;

[0029] In response to the fact that the action gain data of the predicted biomolecule is less than the action gain data of the biomolecule under investigation corresponding to the predicted biomolecule, it is determined that the predicted biomolecule and the action gain data meet a first condition.

[0030] Optionally, the loop submodule is used for:

[0031] A control biomolecule is determined based on the predicted biomolecule generated prior to the predicted biomolecule and the second biomolecule.

[0032] In response to the prediction biomolecule being the same as the control biomolecule, it is determined that the prediction biomolecule and the action gain data meet a first condition.

[0033] Optionally, the second determining module includes:

[0034] The first determining submodule is used to determine the prediction score of each molecule prediction result based on the action gain data.

[0035] The second determination submodule is used to determine the candidate biomolecule from the plurality of molecular prediction results based on the prediction score.

[0036] Optionally, the first determining submodule is used to:

[0037] For each of the molecular prediction results, determine the number of actions required to transform the biomolecule under investigation into the molecular prediction result;

[0038] Based on the action gain data and the number of actions, the prediction score of the molecular prediction result is determined.

[0039] Optionally, determining the prediction score of the molecular prediction result based on the action gain data and the number of actions includes:

[0040] Based on the action gain data, a first score is determined;

[0041] The second score is determined based on the number of actions performed.

[0042] Based on the first score and the second score, the prediction score of the molecular prediction result is determined.

[0043] Optionally, the prediction unit is used for:

[0044] Determine the first molecular sequence information of the first biomolecule and the second molecular sequence information of each of the plurality of candidate biomolecules;

[0045] Extract a first sequence feature from the first molecular sequence information, and extract a second sequence feature from the second molecular sequence information;

[0046] Based on the first sequence features and the second sequence features, the prediction model is used to predict rewards and obtain the molecular docking scores of each of the multiple candidate biomolecules with the first biomolecule.

[0047] Based on the molecular docking score, the reward prediction result for each of the multiple candidate biomolecules is determined.

[0048] Optionally, the prediction model includes a first prediction sub-model, a second prediction sub-model, and a third prediction sub-model;

[0049] The prediction unit is used for:

[0050] For each of the candidate biomolecules, the binding affinity between the candidate biomolecule and the first biomolecule is predicted based on the first prediction sub-model, and a first reward result for the candidate biomolecule is determined based on the binding affinity.

[0051] The solubility of the candidate biomolecule is predicted based on the second prediction sub-model, and a second reward result for the candidate biomolecule is determined based on the solubility.

[0052] The toxicity index of the candidate biomolecule is predicted based on the third prediction sub-model, and the third reward result of the candidate biomolecule is determined based on the toxicity index.

[0053] Based on the first reward result, the second reward result, and the third reward result, the reward prediction result of the candidate biomolecule is determined.

[0054] Optionally, the second biomolecule to be optimized is obtained through the following methods:

[0055] Identify multiple reference biomolecules whose binding affinity to the first biomolecule is greater than a preset threshold;

[0056] Based on the molecular physicochemical information of each of the plurality of reference biomolecules, the second biomolecule is identified among the plurality of reference biomolecules.

[0057] Optionally, the second determining unit is used to:

[0058] Based on the magnitude of the reward prediction results, the multiple candidate biomolecules are arranged in descending order to obtain the molecular sequence;

[0059] The first predetermined number of candidate biomolecules in the molecular sequence are identified as the target biomolecule.

[0060] Optionally, the molecular optimization device further includes an update unit, the update unit being used for:

[0061] Based on preset verification rules, molecular binding verification is performed between the target biomolecule and the first biomolecule to obtain verification results;

[0062] In response to the discrepancy between the verification result and the reward prediction result of the target biomolecule, the prediction model is updated with parameters based on the verification result, the target biomolecule, and the first biomolecule.

[0063] Optionally, the molecular optimization device further includes a training unit, the training unit being used for:

[0064] Multiple sample molecule pairs are obtained, wherein each sample molecule pair includes a first sample biomolecule, a second sample biomolecule, and reference binding data of the first sample biomolecule and the second sample biomolecule;

[0065] First sample sequence features are extracted from the molecular sequence information of the first sample biomolecule, and second sample sequence features are extracted from the molecular sequence information of the second sample biomolecule.

[0066] The first sample sequence features and the second sample sequence features are fused into a sample sequence fusion feature;

[0067] Based on the sample sequence fusion features, the binding affinity of the sample molecule pairs is predicted using the original model to obtain the predicted binding data of the first sample biomolecule and the second sample biomolecule in the sample molecule pairs.

[0068] Based on the reference combined data and the prediction combined data, the original model is trained to obtain the prediction model.

[0069] Optionally, training the original model based on the reference combined data and the prediction combined data to obtain the prediction model includes:

[0070] For each of the sample pairs, a sub-loss function is determined based on the reference binding data and the predicted binding data;

[0071] Based on the sub-loss functions of the multiple sample pairs, the total loss function is determined;

[0072] The original model is trained based on the total loss function to obtain the prediction model.

[0073] According to one aspect of this disclosure, an electronic device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the molecular optimization method as described above.

[0074] According to one aspect of this disclosure, a computer-readable storage medium is provided, the storage medium storing a computer program that, when executed by a processor, implements the molecular optimization method as described above.

[0075] According to one aspect of this disclosure, a computer program product is provided, comprising a computer program that is read and executed by a processor of a computer device, causing the computer device to perform the molecular optimization method as described above.

[0076] In this embodiment, a molecular optimization model is used to perform multiple rounds of optimization predictions on a second biomolecule based on a first biomolecule. This enables the prediction of a large number of molecules with potential optimization value in a short period of time. Simultaneously, the candidate biomolecules in each round are predicted based on the candidate biomolecules from the previous round relative to the current round. This iterative optimization strategy ensures the continuity and cumulative nature of the optimization process. Each round of prediction improves and optimizes upon the previous round, avoiding repetitive predictions from scratch and improving overall optimization efficiency and accuracy. Furthermore, in this embodiment, the molecular optimization model is responsible for the initial optimization prediction, while the prediction model performs reward predictions on the candidate biomolecules. This combination fully leverages the advantages of both models: the molecular optimization model narrows the candidate range through optimization prediction, while the prediction model uses its powerful predictive capabilities to perform reward predictions on the candidate biomolecules. Finally, based on the reward prediction results, candidate biomolecules with better binding affinity to the first biomolecule are selected as target biomolecules, resulting in a significant improvement in binding affinity between the optimized target biomolecule and the original second biomolecule. This approach can improve the accuracy of molecular optimization and generate biomolecules with better binding affinity to the first biomolecule.

[0077] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objectives and other advantages of this disclosure may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0078] The accompanying drawings are provided to further understand the technical solutions of this disclosure and constitute a part of the specification. They are used together with the embodiments of this disclosure to explain the technical solutions of this disclosure and do not constitute a limitation on the technical solutions of this disclosure.

[0079] Figure 1 This is a system architecture diagram of a molecular optimization method applied according to embodiments of the present disclosure;

[0080] Figures 2A-2D A schematic diagram illustrating the application of the molecular optimization method according to embodiments of the present disclosure in a peptide-binding generation scenario is shown.

[0081] Figure 3 This is a flowchart of a molecular optimization method according to an embodiment of the present disclosure;

[0082] Figure 4 This is a flowchart illustrating the determination of a second biomolecule according to an embodiment of this disclosure;

[0083] Figure 5This is a schematic diagram of the molecular structure of a second biomolecule according to an embodiment of the present disclosure;

[0084] Figure 6 This is a flowchart illustrating the generation of multiple candidate biomolecules according to an embodiment of this disclosure;

[0085] Figure 7 This is a flowchart illustrating the generation of predicted biomolecules according to an embodiment of this disclosure;

[0086] Figure 8 This is a flowchart illustrating the determination of motion gain data according to an embodiment of this disclosure;

[0087] Figures 9A-9C This is a schematic diagram illustrating the process of optimizing and predicting a second biomolecule according to an embodiment of this disclosure;

[0088] Figure 10 This is a flowchart illustrating the process of identifying candidate biomolecules according to an embodiment of this disclosure;

[0089] Figure 11 This is a schematic diagram illustrating the overall implementation process of optimizing and predicting a second biomolecule according to an embodiment of the present disclosure;

[0090] Figure 12 This is a flowchart of determining a predicted score according to an embodiment of the present disclosure;

[0091] Figures 13A-13C This is a schematic diagram illustrating the process of determining a predicted score according to an embodiment of the present disclosure;

[0092] Figure 14 This is a flowchart illustrating the determination of reward prediction results according to an embodiment of the present disclosure;

[0093] Figure 15 This is a schematic diagram illustrating the process of determining reward prediction results according to another embodiment of the present disclosure;

[0094] Figure 16 This is a schematic diagram illustrating the process of determining reward prediction results according to an embodiment of the present disclosure;

[0095] Figure 17 This is a flowchart illustrating the determination of a target biomolecule according to an embodiment of the present disclosure;

[0096] Figure 18 This is a schematic diagram illustrating the process of determining a target biomolecule according to an embodiment of the present disclosure;

[0097] Figure 19 This is a flowchart of updating parameters of a prediction model according to an embodiment of the present disclosure;

[0098] Figure 20This is an overall flowchart of training a prediction model according to an embodiment of the present disclosure;

[0099] Figure 21 This is a flowchart of training a prediction model according to another embodiment of the present disclosure;

[0100] Figure 22 These are detailed implementation diagrams of a molecular optimization method according to an embodiment of the present disclosure;

[0101] Figures 23A-23C This is an experimental comparison diagram of the optimized peptide corresponding to the target protein and the bound natural peptide according to an embodiment of the present disclosure;

[0102] Figures 24A-24E This is an experimental comparison diagram of the physicochemical properties of the optimized peptide corresponding to the target protein in one embodiment of this disclosure;

[0103] Figure 25 This is an experimental comparison diagram of multiple iterations in molecular optimization according to an embodiment of the present disclosure;

[0104] Figure 26 This is a block diagram of a molecular optimization apparatus according to an embodiment of the present disclosure;

[0105] Figure 27 This is a terminal structure diagram of a molecular optimization method according to an embodiment of the present disclosure;

[0106] Figure 28 This is a server structure diagram of a molecular optimization method according to an embodiment of the present disclosure. Detailed Implementation

[0107] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this disclosure.

[0108] Before providing a further detailed description of the embodiments of this disclosure, the terms and concepts used in these embodiments are explained, and they are subject to the following interpretations:

[0109] The system architecture and scenarios in which this disclosure is applied are described below.

[0110] Figure 1 This is a system architecture diagram of the molecular optimization method applied according to embodiments of the present disclosure. It includes an object terminal 140, an Internet 130, a gateway 120, a server 110, and a biomolecular database 150, etc.

[0111] The target terminal 140 includes various forms such as desktop computers, laptops, PDAs (personal digital assistants), tablets, mobile phones, in-vehicle terminals, home theater terminals, smart TVs, and dedicated terminals. Furthermore, it can be a single device or a collection of multiple devices. The target terminal 140 can communicate with the Internet 130 via wired or wireless means to exchange data. The target terminal 140 is used to specify a first biomolecule and provide a second biomolecule that needs to be optimized for that biomolecule. It then submits the specified first biomolecule and the second biomolecule to be optimized to a server, so that the server can optimize the second biomolecule into a target biomolecule with a better binding affinity to the first biomolecule, based on the first biomolecule.

[0112] Server 110 refers to a computer system that can provide certain services to object terminal 140. Compared to ordinary object terminal 140, server 110 has higher requirements in terms of stability, security, and performance. Server 110 can be a high-performance computer in a network platform, a cluster of multiple high-performance computers, a portion of a high-performance computer (e.g., a virtual machine), a combination of portions of multiple high-performance computers (e.g., virtual machines), or a cloud server, etc. Server 110 contains various types of services, and the implementation of each service of server 110 is often associated with some intermediate databases or storage media. Server 110 is used to call a trained molecular optimization model to optimize and predict a second biomolecule based on a first biomolecule specified by the object, generate multiple candidate biomolecules, call the trained prediction model to determine the reward score of each of the multiple candidate biomolecules, and determine the final optimized target biomolecule from the multiple candidate biomolecules based on the reward score. Biomolecule database 150 is used to store multiple existing biomolecules and multiple generated candidate biomolecules. Database 150 can be set up separately or integrated into video processing server 110 or other electronic devices.

[0113] Gateway 120, also known as an internetwork connector or protocol converter, is a computer system or device that acts as a translator, enabling network interconnection at the transport layer. It bridges the gap between two systems using different communication protocols, data formats, languages, or even completely different architectures. Gateways can also provide filtering and security functions. Messages sent from target terminal 140 to server 110 are forwarded to the corresponding server via gateway 120. Messages sent from server 110 to target terminal 140 are also forwarded to the corresponding target terminal 140 via gateway 120.

[0114] The embodiments disclosed herein can be applied in various scenarios, such as Figures 2A-2D The examples shown include scenarios involving the generation of binding peptides.

[0115] like Figure 2A As shown, when an object wants to generate a binding peptide corresponding to a specific target protein, it executes the binding peptide generation process through the molecular processing platform on the object's terminal. At this time, the molecular processing platform page displays a prompt field "Please specify the first biomolecule and the second biomolecule to be optimized, etc.", and provides editing areas for inputting the first biomolecule, the second biomolecule, and setting the iteration rounds. Based on this, the object enters "target protein K" in the editing area for inputting the first biomolecule, "natural peptide M" in the editing area for inputting the second biomolecule, and "6" in the editing area for setting the iteration rounds, and clicks the "OK" button to confirm that the natural peptide M will be optimized into a binding peptide with good binding affinity to target protein K.

[0116] like Figure 2B As shown, after the object clicks the "OK" button, a prompt window will appear on the page. The prompt window has the prompt field "Using molecular optimization model to optimize and predict the second biomolecule, please wait patiently...", to inform the object that the second biomolecule is being optimized and predicted, and what kind of biomolecule can be transformed from the second biomolecule.

[0117] like Figure 2C As shown, once the optimization prediction for the second biomolecule is complete, the page will display the prompt field "The following are the candidate binding peptides predicted in each iteration round. Reward scoring is being performed on each candidate binding peptide," and will show the candidate binding peptides for each iteration round, along with their peptide sequence information. Specifically, the peptide sequence information for candidate binding peptide 1 generated in iteration round 1 is ASPSSL; the peptide sequence information for candidate binding peptide 2 generated in iteration round 2 is ASASSL; the peptide sequence information for candidate binding peptide 3 generated in iteration round 3 is ASASSM; the peptide sequence information for candidate binding peptide 4 generated in iteration round 4 is KSASSM; the peptide sequence information for candidate binding peptide 5 generated in iteration round 5 is KTASSM; and the peptide sequence information for candidate binding peptide 6 generated in iteration round 6 is LTASSM.

[0118] like Figure 2DAs shown, after the reward scoring for each candidate binding peptide is completed, the page will display a prompt field: "The specific details of the reward scoring for each candidate binding peptide are as follows; the top three candidate binding peptides with the highest reward scores will be selected as the final optimization result." Specifically, since candidate binding peptide 1 has a reward score of 75; candidate binding peptide 2 has a reward score of 83; candidate binding peptide 3 has a reward score of 78; candidate binding peptide 4 has a reward score of 85; candidate binding peptide 5 has a reward score of 88; and candidate binding peptide 6 has a reward score of 92, candidate binding peptide 4, candidate binding peptide 5, and candidate binding peptide 6 will be selected as the final optimization result.

[0119] The embodiments of this disclosure are described in general below.

[0120] According to one embodiment of this disclosure, a molecular optimization method is provided.

[0121] This molecular optimization method is generally applied in business scenarios where it is necessary to generate biomolecules that can bind relatively stably to specific target proteins or antigens, for example... Figures 2A-2D The examples shown depict the peptide generation scenario. This disclosure provides a molecular optimization scheme based on deep reinforcement learning, which can improve the accuracy of molecular optimization and thus generate biomolecules with better binding affinity to the first biomolecule.

[0122] like Figure 3 As shown, the molecular optimization method according to an embodiment of this disclosure can be executed by an electronic device, which may be... Figure 1 The server or object terminal shown. A molecular optimization method according to one embodiment of this disclosure may include:

[0123] Step 310: Determine the first biomolecule and the second biomolecule to be optimized;

[0124] Step 320: Based on the first biomolecule, optimize and predict the second biomolecule using a preset molecular optimization model to determine the candidate biomolecules predicted in each round.

[0125] Step 330: Based on the pre-trained prediction model, perform reward prediction on multiple candidate biomolecules to obtain the reward prediction results for each candidate biomolecule.

[0126] Step 340: Based on the reward prediction results, identify the target biomolecule from multiple candidate biomolecules.

[0127] Steps 310-340 are described in detail below.

[0128] In step 310, a first biomolecule and a second biomolecule to be optimized are determined.

[0129] The first biomolecule refers to biomolecules such as specific target proteins and antigens.

[0130] The second biomolecule refers to a biomolecule that has been optimized to bind with the first biomolecule. For example, the second biomolecule can be a binding peptide corresponding to a specific target protein, an antibody corresponding to an antigen, and so on.

[0131] In this specific implementation, during molecular optimization, the first biomolecule is often specified by the object. Based on the molecule name or molecule number input by the object at the object terminal, the first biomolecule can be determined from a preset biomolecule database. Then, for the second biomolecule to be optimized, a biomolecule different from the first biomolecule can be arbitrarily selected from the biomolecule database as the second biomolecule.

[0132] In step 320, based on the first biomolecule, the second biomolecule is optimized and predicted using a preset molecular optimization model to determine the candidate biomolecules predicted in each round.

[0133] A molecular optimization model refers to a neural network model that can optimize and predict the second biomolecule based on the molecular physicochemical information of the first biomolecule and the molecular physicochemical information of the second biomolecule. This molecular optimization model can be a deep learning model, which can predict the possible action changes that may be made to the biomolecule based on the current biomolecule.

[0134] Optimization prediction refers to predicting possible action transformations for a current biomolecule based on a molecular optimization model, and then examining the differences in molecular properties between the biomolecule generated by the action transformation and the current biomolecule after the prediction based on the molecular optimization model, in order to predict action transformations that can generate biomolecules with better performance as much as possible.

[0135] In this embodiment of the disclosure, the molecular optimization model can be a neural network model based on fully connected neural networks, convolutional neural networks, Transformer networks, etc., without any specific limitations.

[0136] Rounds are used to indicate the various iterations in which the second biomolecule is optimized for prediction.

[0137] Candidate biomolecules refer to biomolecules with superior molecular performance generated in each round of iterative prediction.

[0138] In this round, the candidate biomolecules are predicted based on the candidate biomolecules of the previous round.

[0139] To save space, the specific process of optimizing and predicting the second biomolecule based on the first biomolecule using a preset molecular optimization model, as described in this embodiment, will be described in detail below, and will not be repeated here.

[0140] In step 330, reward predictions are performed on multiple candidate biomolecules based on a pre-trained prediction model to obtain the reward prediction results for each candidate biomolecule.

[0141] A predictive model is a neural network model that can assign a reward score to a candidate biomolecule based on its molecular properties or binding affinity to a primary biomolecule. Predictive models can be affinity prediction models, solubility prediction models, or toxicity prediction models, among others.

[0142] For example, in order to improve the affinity between candidate biomolecules and the first biomolecule, an affinity prediction model can be used as the prediction model.

[0143] The reward prediction result is used to indicate the quality of candidate biomolecules in terms of molecular performance, binding affinity with the first biomolecule, and other indicators in a fractional form. A higher reward prediction result indicates better performance of the candidate biomolecule.

[0144] To save space, the specific process of reward prediction for multiple candidate biomolecules based on the prediction model in this embodiment will be described in detail below, and will not be repeated here.

[0145] In step 340, the target biomolecule is identified from multiple candidate biomolecules based on the reward prediction results.

[0146] The target biomolecule is used to indicate the final optimized candidate biomolecule with better molecular performance and binding affinity to the first biomolecule.

[0147] Among them, the binding affinity between the target biomolecule and the first biomolecule is greater than the binding affinity between the second biomolecule and the first biomolecule.

[0148] In this specific implementation, a larger reward prediction result indicates better performance of the candidate biomolecule. Based on this, several candidate biomolecules with larger reward prediction results can be identified as the target biomolecules.

[0149] Through steps 310-340 above, in this embodiment of the disclosure, a molecular optimization model is used to perform multiple rounds of optimization prediction on a second biomolecule based on a first biomolecule. This enables the prediction of a large number of molecules with potential optimization value in a short period of time. Simultaneously, the candidate biomolecules in each round are predicted based on the candidate biomolecules of the previous round relative to the current round. This iterative optimization strategy ensures the continuity and cumulative nature of the optimization process. Each round of prediction improves and optimizes upon the previous round, avoiding repetitive predictions from scratch and improving overall optimization efficiency and accuracy. Furthermore, in this embodiment of the disclosure, the molecular optimization model is responsible for the initial optimization prediction, while the prediction model performs reward prediction on the candidate biomolecules. This combination fully leverages the advantages of both models. The molecular optimization model narrows the candidate range through optimization prediction, while the prediction model uses its powerful predictive capabilities to perform reward prediction on the candidate biomolecules. Finally, based on the reward prediction results, candidate biomolecules with better binding affinity to the first biomolecule are selected as target biomolecules, resulting in a significant improvement in binding affinity between the optimized target biomolecule and the original second biomolecule. This approach can improve the accuracy of molecular optimization and generate biomolecules with better binding affinity to the first biomolecule.

[0150] The above is a general description of steps 310-340. The following will provide a detailed description of the specific implementation of steps 310-340.

[0151] Step 310 will be described in detail below.

[0152] In step 330, a first biomolecule and a second biomolecule to be optimized are determined.

[0153] In a specific implementation of this embodiment, a biomolecule can be randomly extracted from a preset biomolecule database as a second biomolecule.

[0154] For example, when the first biomolecule is a specific target protein, a polypeptide sequence can be randomly extracted from a biomolecule database as the second biomolecule.

[0155] The initial selection of the second biomolecule significantly impacts molecular optimization efficiency. For example, if the initially selected second biomolecule has poor molecular properties, it often requires more time and resources to optimize it into a better-performing biomolecule. However, if the initially selected second biomolecule already has good molecular properties, there is a greater chance of optimizing it into an even better-performing biomolecule. Therefore, this disclosure provides a scheme for determining the second biomolecule based on its molecular properties, which can effectively improve molecular optimization efficiency.

[0156] Please refer to Figure 4 In one embodiment, the second biomolecule to be optimized is obtained by:

[0157] Step 410: Identify multiple reference biomolecules whose binding affinity to the first biomolecule is greater than a preset threshold;

[0158] Step 420: Based on the molecular physicochemical information of multiple reference biomolecules, identify the second biomolecule among the multiple reference biomolecules.

[0159] Steps 410-420 are described in detail below.

[0160] In step 410, binding affinity is used to indicate the interaction between the first biomolecule and the reference biomolecule. The reference biomolecule refers to a biomolecule that can bind to the first biomolecule and is available prior to performing the molecular optimization method.

[0161] In this specific implementation, firstly, for each biomolecule in the preset biomolecule database, the binding affinity of each biomolecule to the first biomolecule is determined. Then, biomolecules with a binding affinity to the first biomolecule greater than a preset threshold are identified as reference biomolecules.

[0162] In step 420, molecular physicochemical information is used to indicate the physicochemical properties of a reference biomolecule, wherein the physicochemical properties include, but are not limited to, the aromaticity, charge, isoelectric point, global hydrophobicity, global hydrophobic moment, etc. of the reference biomolecule.

[0163] It should be noted that the physicochemical properties of the reference biomolecule often affect the interaction between the reference biomolecule and the first biomolecule, as well as its distribution and metabolism within the organism.

[0164] In this specific implementation, firstly, for each reference biomolecule, a molecular score is calculated based on its molecular physicochemical information to obtain the molecular physicochemical score of the reference biomolecule. Then, among the multiple reference biomolecules, the reference biomolecule with the highest molecular physicochemical score is selected as the second biomolecule.

[0165] In another embodiment, a second biomolecule may be randomly selected from a plurality of reference biomolecules whose molecular physicochemical fraction is greater than a preset fraction threshold.

[0166] Furthermore, when scoring the reference biomolecule based on its molecular physicochemical information, this information often contains multiple types of physicochemical information items; for example, aromaticity and charge are different types of physicochemical information items. Therefore, firstly, a scoring function is used to score each physicochemical information item in the molecular physicochemical information, obtaining a physicochemical score for each item. Then, the physicochemical scores of multiple items are weighted and calculated to obtain the molecular physicochemical score of the reference biomolecule.

[0167] For example, when the first biomolecule is a specific target protein, a known binding natural peptide corresponding to that target protein is extracted from a biomolecule database as the second biomolecule.

[0168] like Figure 5 The diagram shows a schematic representation of the molecular sequence matrix of a polypeptide as the second biomolecule. "20" represents the 20 standard amino acid types found in nature, and "L" represents the number of positions in the polypeptide sequence. L is a positive integer. Specifically, Figure 5 The second biomolecule is a polypeptide sequence of length L. In the molecular sequence matrix, each column corresponds to a specific position in the polypeptide sequence, and the row represents the amino acid type at that position. For each column, only the element pointed to by the amino acid type at that position has a value of 1, and the element values ​​of all other elements are 0. Specifically, the amino acid type at the first sequence position of the polypeptide sequence is A; the amino acid type at the second sequence position is G; the amino acid type at the third sequence position is D; the amino acid type at the fourth sequence position is D; the amino acid type at the fifth sequence position is A; ...; and the amino acid type at the Lth sequence position is Y. Based on this, the second biomolecule can be represented in sequence form as [AGDDA…Y].

[0169] The advantage of this embodiment is that, in determining the second biomolecule, it takes into account multiple methods such as random selection and molecular screening based on binding affinity, molecular physicochemical information, etc., which improves the flexibility of determining the second biomolecule to a certain extent and can meet the molecular optimization needs in various business scenarios.

[0170] Step 320 will be described in detail below.

[0171] In step 320, based on the first biomolecule, the second biomolecule is optimized and predicted using a preset molecular optimization model to determine the candidate biomolecules predicted in each round, wherein the binding affinity between the target biomolecule and the first biomolecule is greater than the binding affinity between the second biomolecule and the first biomolecule.

[0172] Please refer to Figure 6 In one embodiment, step 320 specifically includes, but is not limited to, the following steps 610-640:

[0173] Step 610: Initialize the current round number to 1, and initialize the biomolecule to be examined to the second biomolecule;

[0174] Step 620: In the current round, based on the first biomolecule, optimize and predict the biomolecule under investigation using a molecular optimization model, and determine multiple molecular prediction results and the action gain data of each of the multiple molecular prediction results.

[0175] Step 630: Based on action gain data, candidate biomolecules are identified from multiple molecular prediction results;

[0176] Step 640: Increment the round number by 1, update the biomolecule to be examined to the candidate biomolecule, and return to the current round. Based on the first biomolecule, optimize and predict the biomolecule to be examined using the molecular optimization model, determine multiple molecular prediction results and the action gain data of each of the multiple molecular prediction results, until the round number is the target number, and obtain multiple candidate biomolecules.

[0177] Steps 610-640 are described in detail below.

[0178] In step 610, the current round refers to the round corresponding to the current optimization prediction, the round number is used to indicate the number of the current round, and the biomolecule to be examined refers to the specific biomolecule to be optimized and predicted.

[0179] In this specific implementation, the optimization prediction for the second biomolecule is performed iteratively round by round, and each round involves further optimization prediction of the candidate biomolecules predicted in the previous round. Therefore, starting from the first round, the round number is initialized to 1 to indicate that the optimization prediction for the first round is being executed. Simultaneously, the biomolecule to be examined at this point is initialized as the second biomolecule.

[0180] In step 620, the molecular prediction results are used to indicate what the molecular sequence of the predicted biomolecule is.

[0181] Action gain data is used to indicate the magnitude of the value of the action transformation from the second biomolecule transformation to the molecular prediction result.

[0182] In this specific implementation, the optimization search for the biomolecule under investigation in the current round can be based on the Monte Carlo tree search algorithm. Specifically, a molecular optimization model is used to simulate the environment based on the molecular information of the first biomolecule and the biomolecule under investigation. In the current round, the initial biomolecule under investigation is taken as the root node. Starting from the root node of the tree, the search is continuously and progressively downwards through simulation until an unpredicted molecular sequence is reached as the molecular prediction result.

[0183] To save space, the specific process of optimizing and predicting the biomolecule under investigation based on the first biomolecule using a molecular optimization model in the embodiments of this disclosure will be described in detail below, and will not be repeated here.

[0184] In step 630, since a larger action gain data indicates a higher value of the action transformation from the second biomolecule to the molecular prediction result, this action transformation is more consistent with the actual expectation. Based on this, the action gain data of multiple molecular prediction results can be compared, and the molecular prediction result with the largest action gain data can be used as the candidate biomolecule for the current round.

[0185] In step 640, the target number is used to indicate the total number of rounds of optimized predictions pre-set for the second biomolecule.

[0186] In this specific implementation, after the first round of optimization prediction is completed, the second round of optimization prediction needs to begin. At this time, the round number is incremented by 1, making the current round number 2, indicating that the second round of optimization prediction is in progress. Since the candidate biomolecules predicted in the first round often have better molecular properties than the initial second biomolecule, to improve the accuracy of the optimization prediction, the biomolecule to be examined is updated to a candidate biomolecule. This allows for the optimization prediction of the candidate biomolecules generated in the first round in the second round. The specific prediction method in the second round is consistent with step 620 above. This process continues until the current round is the last round. At this point, the round number of the current round is the target number. After completing the optimization prediction of the current round, the entire optimization prediction process is confirmed to be complete, and the candidate biomolecules generated in each round are obtained.

[0187] The advantage of this embodiment is that, through each iteration, the molecular optimization model is used to gradually optimize the biomolecule under investigation, making the optimized prediction for the second biomolecule closer to the target characteristics or functions. Each round is based on the candidate biomolecules of the previous round for optimization prediction. As the number of rounds increases, the optimization effect of each round will continue to accumulate, and the final candidate biomolecule will have better performance and characteristics. This approach can improve prediction efficiency and prediction accuracy.

[0188] Please refer to Figure 7 In one embodiment, step 620 specifically includes, but is not limited to, the following steps 710-730:

[0189] Step 710: In the current round, extract the molecular feature information of the first biomolecule;

[0190] Step 720: Based on molecular feature information, optimize and predict the biomolecule under investigation using a molecular optimization model to obtain the predicted biomolecule, and determine the action gain data of the predicted biomolecule based on the difference between the predicted biomolecule and the second biomolecule.

[0191] Step 730: Update the biomolecule to be examined with the predicted biomolecule, and return the steps of optimizing the prediction of the biomolecule to be examined through the molecular optimization model based on molecular feature information, until the predicted biomolecule and the action gain data meet the first condition, the predicted biomolecule is determined as the molecular prediction result, and the action gain data of the predicted biomolecule is determined as the action gain data of the molecular prediction result.

[0192] Steps 710-730 are described in detail below.

[0193] In step 710, molecular feature information is used to indicate the three-dimensional structure, molecular sequence and other feature information of the first biomolecule.

[0194] In the specific implementation of this embodiment, in the current round, graph neural networks or attention mechanisms can be used to extract the atomic type, bond type, chemical properties, molecular shape and other feature information of the first biomolecule from its structure, thereby obtaining the molecular feature information of the first biomolecule.

[0195] In step 720, the predicted biomolecule refers to the biomolecule obtained by the molecular optimization model through action transformations of elements at various sequence positions of the biomolecule under investigation.

[0196] In this specific implementation, firstly, the molecular feature information of the first biomolecule and the molecular feature information of the biomolecule to be examined are both input into the molecular optimization model. Next, the molecular optimization model performs feature extraction on the molecular feature information of the first biomolecule and the biomolecule to be examined, extracting a first molecular feature from the first biomolecule's molecular feature information and a second molecular feature from the biomolecule to be examined's molecular feature information. Then, the molecular optimization model performs feature transformation on the second molecular feature, using the first molecular feature as a constraint, to represent the action transformation that can be performed on this biomolecule to be examined. The biomolecule obtained based on this action transformation is used as the predicted biomolecule.

[0197] Furthermore, when the molecular optimization model performs feature transformation on the second molecular feature under the conditional constraints of the first molecular feature, it linearly projects the first molecular feature through the bond channel to obtain the bond vector, and then linearly projects it through the value channel to obtain the value channel. Simultaneously, the molecular optimization model linearly projects the second molecular feature through the query channel to obtain the query vector. Next, attention is calculated on the query vector, bond vector, and value vector to obtain the attention vector, and action prediction is performed based on the attention vector to achieve feature transformation of the second molecular feature and obtain the action transformations that the biomolecule under investigation can perform.

[0198] In step 730, the first condition is used to define the specific circumstances that trigger the termination of the optimization prediction in each round.

[0199] In the specific implementation of this embodiment, in order to achieve deeper molecular optimization prediction in the current round, the predicted biomolecules are updated with the predicted biomolecules to be examined, and the steps of optimizing the predicted biomolecules to be examined through the molecular optimization model based on molecular feature information are returned to obtain the predicted biomolecules. This is to achieve optimized prediction for the predicted biomolecules predicted in the previous step. This process is repeated to continuously increase the prediction depth and achieve multiple iterative predictions in each round, so as to discover more predicted biomolecules in each round.

[0200] Furthermore, when all generated predicted biomolecules and their action gain data meet the first condition, the loop execution of the optimization prediction stops, and the current round of optimization prediction is terminated. Then, all predicted biomolecules generated in the current round are designated as molecular prediction results, and the action gain data of each predicted biomolecule is designated as the action gain data of the molecular prediction results.

[0201] To save space, the specific process of determining whether the predicted biomolecules and action gain data meet the first condition according to the embodiments of this disclosure will be described in detail below. It will not be repeated here.

[0202] The advantages of this embodiment are that by using a molecular optimization model to evaluate and screen multiple predicted biomolecules in each round, this approach can efficiently predict the chemical space of each predicted biomolecule. Simultaneously, by calculating and analyzing the action gain data of the predicted biomolecules, the potential value and improvement space of each predicted biomolecule in the optimization process can be quantified, providing a more reliable basis for selecting candidate biomolecules. Furthermore, considering the setting of termination conditions for each round of iterative prediction, the depth of optimization prediction in each round can be better controlled to a certain extent.

[0203] Please refer to Figure 8In one embodiment, step 720 specifically includes, but is not limited to, the following steps 810-820:

[0204] Step 810: Based on the difference between the predicted biomolecule and the second biomolecule, determine the action space matrix for the transformation from the second biomolecule to the predicted biomolecule;

[0205] Step 820: Based on the action space matrix and molecular feature information, gain prediction is performed using a molecular optimization model to obtain action gain data.

[0206] Steps 810-820 are described in detail below.

[0207] In step 810, the action space matrix is ​​used to indicate what kind of action transformation is performed from the second biomolecule to the predicted biomolecule. The action space matrix can characterize the sequence position pointed to by the action transformation and what element changes have occurred at that sequence position in matrix form.

[0208] In this specific implementation, firstly, the predicted biomolecule and the second biomolecule are converted into matrix representations, resulting in a molecular sequence matrix corresponding to the predicted biomolecule and a molecular sequence matrix corresponding to the second biomolecule, both of which have the same size. Next, the elements in the same columns of both the predicted and second biomolecule molecular sequence matrices are compared, and the columns with different elements are designated as target columns. Further, when determining the action space matrix, its size is set to be the same as the molecular sequence matrix corresponding to the predicted biomolecule. The element values ​​in the columns of the action space matrix with column numbers matching the target column are set to be consistent with the target column's values. All element values ​​in the other columns of the action space matrix, except for the target column, are set to 0, thus obtaining the action space matrix.

[0209] In step 820, the action space matrix and molecular feature information are input into the molecular optimization model. The molecular optimization model aligns and integrates the action sequences corresponding to the action space matrix and the molecular feature sequences corresponding to the molecular feature information to obtain an integration result. Based on the integration result, the degree of improvement relative to the second biomolecule after the action is executed is predicted, and action gain data is output. The molecular optimization model can use linear equations, decision trees, or Gaussian process algorithms to predict the degree of improvement relative to the second biomolecule after the action is executed based on the integration result.

[0210] The advantage of this embodiment is that by analyzing the differences between the predicted biomolecule and the second biomolecule, the action space matrix for the transformation from the second biomolecule to the predicted biomolecule can be precisely defined. This precise definition of the action space helps to clarify the specific chemical structural features and change paths that need to be focused on and adjusted during the optimization process. Furthermore, by utilizing the action space matrix and molecular feature information, gain prediction through a molecular optimization model can efficiently evaluate the potential contribution and effect of different action transformations on biomolecule optimization. This gain prediction mechanism makes the optimization process more purposeful and efficient, enabling the rapid screening of actions with high optimization potential. This approach, which combines the action space matrix and gain prediction, allows for more reasonable predictive decisions at each optimization step, thereby improving the performance and properties of biomolecules and contributing to obtaining high-quality, high-reliability optimization results that can meet the molecular optimization needs of specific biological functions or drug properties.

[0211] In one embodiment, the specific process of determining in step 730 that the predicted biomolecules and action gain data meet the first condition includes, but is not limited to, the following steps:

[0212] Determine the action gain data of the biomolecule under investigation corresponding to the predicted biomolecule;

[0213] In response to the fact that the action gain data of the predicted biomolecules are all less than the action gain data of the biomolecule under investigation corresponding to the predicted biomolecules, it is determined that the predicted biomolecules and the action gain data meet the first condition.

[0214] Specifically, when determining the action gain data of the biomolecule to be examined corresponding to the predicted biomolecule, since the biomolecule to be examined refers to the predicted biomolecule generated in the previous iteration round, the method for determining the action gain data of the biomolecule to be examined corresponding to the predicted biomolecule is similar to steps 810-820 above. To save space, it will not be described in detail again.

[0215] Furthermore, for each predicted biomolecule, the action gain data of the predicted biomolecule is compared with the action gain data of its corresponding biomolecule under investigation. If the action gain data of a predicted biomolecule is greater than or equal to the action gain data of its corresponding biomolecule under investigation, it indicates that the molecular performance of the predicted biomolecule is superior to that of its corresponding biomolecule under investigation. Further optimization prediction for this predicted biomolecule may result in the prediction of a biomolecule superior to the predicted biomolecule, so molecular optimization prediction can continue according to step 730. However, when the action gain data of all predicted biomolecules is less than the action gain data of their corresponding biomolecules under investigation, it indicates that the molecular performance of the predicted biomolecules is inferior to that of their corresponding biomolecules under investigation. Further optimization prediction for this predicted biomolecule is likely to lead to a deterioration in the molecular performance of the generated biomolecule. In this case, the predicted biomolecule and its action gain data meet the first condition, and the current round of optimization prediction is terminated.

[0216] The advantage of this embodiment is that, when controlling the termination of optimization prediction in each round, it takes into account the possibility that as the depth of optimization prediction increases, the action gain of newly generated predicted biomolecules may be lower than that of the previously generated predicted biomolecules. By setting the first condition to terminate prediction when the action gain data of newly generated predicted biomolecules is smaller than that of the previously generated predicted biomolecules, the possibility of executing actions that degrade the performance of predicted biomolecules can be reduced, and the optimization prediction can be made to develop in the direction of predicting biomolecules with better molecular performance as much as possible, thereby improving the effectiveness and accuracy of molecular optimization prediction.

[0217] In another embodiment, the specific process of determining in step 730 that the predicted biomolecules and action gain data meet the first condition includes, but is not limited to, the following steps:

[0218] A control biomolecule is determined based on the predicted biomolecule and the second biomolecule generated prior to the prediction biomolecule.

[0219] In response to the prediction biomolecule being identical to the control biomolecule, the prediction biomolecule and action gain data are determined to meet the first condition.

[0220] Among them, the control biomolecules refer to all the predictive biomolecules and the second biomolecules generated before the predictive biomolecules.

[0221] Specifically, firstly, all predicted biomolecules and second biomolecules generated before the predicted biomolecule are designated as control biomolecules. Next, the predicted biomolecule is compared with each control biomolecule. If the predicted biomolecule is different from each control biomolecule, it indicates that the predicted biomolecule is a new biomolecule that has not appeared before. Continuing to optimize the prediction for this predicted biomolecule may lead to the prediction of other new biomolecules, so molecular optimization prediction can continue according to step 730. However, if the predicted biomolecule is the same as a control biomolecule, it indicates that the current action change is invalid; the predicted biomolecule has become a previously generated biomolecule, indicating a loop or repetition. Continuing to optimize the prediction for this predicted biomolecule will not bring new information or improvement. At this point, the predicted biomolecule and action gain data are determined to meet the first condition, and the current round of optimization prediction is terminated.

[0222] The advantage of this embodiment is that, when controlling the termination of optimization prediction in each round, it takes into account the possibility that as the depth of optimization prediction increases, newly generated predicted biomolecules may be previously generated predicted biomolecules. The first condition is set to terminate the prediction once the newly generated predicted biomolecule is the same as all predicted biomolecules generated before the predicted biomolecule and the second biomolecule. This can reduce the waste of resources caused by repeated prediction of the same predicted biomolecule, avoid optimization prediction from getting stuck in a local loop, and make optimization prediction develop as much as possible in the direction of predicting biomolecules that have never appeared before, thereby improving the effectiveness and accuracy of molecular optimization prediction.

[0223] The following is combined Figures 9A-9C The specific process of molecular optimization prediction using molecular optimization models is described with examples.

[0224] like Figure 9AAs shown, for a second biomolecule with the molecular sequence [AGDDAC], a predicted biomolecule with the molecular sequence [ADDDAC] was generated. Based on this, firstly, both the second biomolecule and the predicted biomolecule are converted into matrix form. Both the molecular sequence matrix corresponding to the second biomolecule and the molecular sequence matrix corresponding to the predicted biomolecule are 6×20 matrices, with sequence positions as columns and different amino acid types as rows. Next, by comparing the molecular sequence matrices corresponding to the second biomolecule and the predicted biomolecule, it is determined that the elements with a value of 1 in the second column of the two molecular sequence matrices are different. When transforming from the second biomolecule to the predicted biomolecule, the element with a value of 1 in the second column changes from corresponding to amino acid type G to corresponding to amino acid type D. Based on this, it can be determined that the action space matrix is ​​a 6×20 matrix where the element corresponding to amino acid type D in the second column has a value of 1, and all other elements have a value of 0.

[0225] like Figure 9B The diagram illustrates the specific optimization prediction process of the molecular optimization model for the biomolecule under investigation. Specifically, the molecular optimization model is a neural network model consisting of an input layer, hidden layers, and an output layer. First, the biomolecule under investigation is input into the molecular optimization model. The input layer extracts features from the biomolecule, the hidden layer aggregates the extracted molecular features, and the output layer predicts the possible action transformations for the biomolecule under investigation, the value of each action transformation (action gain data V), and the probability of selecting an action transformation (action probability data P) based on the aggregated feature vector.

[0226] like Figure 9CThe diagram illustrates the entire process of a molecular optimization model predicting and optimizing a second biomolecule. Specifically, the state space of the second biomolecule to be optimized is denoted as state0, which is equivalent to a molecular sequence matrix. First, for the second biomolecule, under the influence of the molecular optimization model on the environmental state, action 1 transforms the second biomolecule into the first predicted biomolecule, and the state space of the first predicted molecule is denoted as state1. Next, for the first predicted biomolecule, under the influence of the molecular optimization model on the environmental state, action 2 transforms this predicted biomolecule into a new predicted biomolecule, and the state space of the second predicted molecule is denoted as state2. This process continues until action k transforms the predicted biomolecule into a state with state k, at which point the prediction terminates. The state spaces state0, state1, state2, and statek are all distinct. The termination condition for this optimization prediction round is that the action gain data of the k-th predicted biomolecule is less than that of the previous predicted biomolecule, or the current iteration count reaches the preset iteration count, or the current action k is invalid, where k is a positive integer.

[0227] Please refer to Figure 10 In one embodiment, step 630 specifically includes, but is not limited to, the following steps 1010-1020:

[0228] Step 1010: For each molecule prediction result, determine the prediction score of the molecule prediction result based on the action gain data;

[0229] Step 1020: Based on the prediction scores, candidate biomolecules are identified from multiple molecular prediction results.

[0230] Steps 1010-1020 are described in detail below.

[0231] In step 1010, the prediction score is used to indicate the result of value scoring of the predicted molecular prediction results.

[0232] In this specific implementation, since the motion gain data is numerical data, the prediction score can be determined based on the motion gain data using either a lookup table or a function calculation method. Specifically, when using a lookup table, firstly, a preset mapping table is obtained, where the mapping table indicates the candidate scores corresponding to each preset motion gain interval. Next, for the molecular prediction result, the motion gain interval in the mapping table where the motion gain data is located is determined, and the candidate score corresponding to that interval is determined as the prediction score of the molecular prediction result. When using a function calculation method, firstly, a preset scoring function is called, where the scoring function is an increasing function with the motion gain data as the independent variable and the prediction score as the dependent variable. Next, the motion gain data of the molecular prediction result is input into the scoring function, and the output of the scoring function is used as the prediction score of the molecular prediction result.

[0233] In step 1020, the molecule with the highest prediction score can be selected as a candidate biomolecule based on the prediction score.

[0234] like Figure 11The diagram illustrates the optimization prediction of a second biomolecule S0 with the molecular sequence [ASPMSL]. Specifically, firstly, for the second biomolecule S0 with the molecular sequence [ASPMSL], a molecular optimization model is used to optimize and predict its molecular sequence, resulting in two predictions: [ASFMSL] and [ASPSSL]. [ASFMSL] is formed by changing the amino acid at the third sequence position of the second biomolecule to F, while [ASPSSL] is formed by changing the amino acid at the fourth sequence position to S. Next, for the prediction [ASPSSL], the molecular optimization model is used to optimize and predict its molecular sequence, resulting in two new predictions: [AEPSSL] and [ASPSSM]. Simultaneously, for the prediction [ASFMSL], the molecular optimization model is also used to optimize and predict its molecular sequence, but no new molecular sequence is predicted. Similarly, the molecular prediction results [AEPSSL] and [ASPSSM] can be further optimized to predict more new molecular sequences until the round starting with the second biomolecule (the horizontally expanded prediction chain) satisfies the first condition mentioned above. Further, for the second biomolecule S0 with the molecular sequence [ASPMSL], the molecular prediction results [ASFMSL], [ASPSSL], [AEPSSL], and [ASPSSM] belonging to the same round, prediction scores are calculated, and the molecular prediction result [ASPSSL] with the highest prediction score is selected as candidate biomolecule S1. Further, in the next round, a new horizontally expanded prediction is performed starting with candidate biomolecule S1, and the above process is repeated. The molecular prediction result [ASASSL] with the highest prediction score in this round is identified as candidate biomolecule S2, and so on, until a candidate biomolecule Sn with the molecular sequence [LTASSM] is identified in the nth round. Only then is the optimized prediction of the second biomolecule S0 considered complete.

[0235] The advantage of this embodiment is that, since action gain data can indicate the action value of transforming the biomolecule under investigation into the molecule prediction result, the greater the action value, the more conducive the action transformation is to optimizing for a better biomolecule. Determining the prediction score based on action gain data enables the accurate quantification of the action value of each molecule's prediction result using the score, and the candidate biomolecules can also be identified more intuitively and accurately based on the prediction score.

[0236] Please refer to Figure 12In one embodiment, step 1010 specifically includes, but is not limited to, the following steps 1210-1220:

[0237] Step 1210: For each molecule prediction result, determine the number of actions required to transform the biomolecule under investigation into the molecule prediction result;

[0238] Step 1220: Determine the prediction score of the molecular prediction result based on the action gain data and the number of actions.

[0239] Steps 1210-1220 are described in detail below.

[0240] In step 1210, the action count indicates the total number of action transformations required to transform the molecular sequence of the biomolecule under investigation into the exact same molecular sequence as the predicted molecular sequence during the optimization prediction. Typically, each action transformation only affects one element of the molecular sequence of the biomolecule under investigation.

[0241] In this specific implementation, since each action transformation in the optimization prediction of the biomolecule under investigation can be continuously accumulated by a counter, with authorization, the accumulated value corresponding to the transformation from the biomolecule under investigation to the molecular prediction result can be read from the counter, and the read accumulated value is used as the number of actions from the biomolecule under investigation to the molecular prediction result.

[0242] For example, Figure 11 In this study, the molecular prediction results [AEPSSL] and [ASPSSM] can be considered as obtained by the second biomolecule through two action transformations, with a total of 2 actions. The molecular prediction results [ASFMSL] and [ASPSSL] can be considered as obtained by the second biomolecule through one action transformation, with a total of 1 action.

[0243] In step 1220, the specific process of determining the predicted score based on the action gain data and the number of actions may include, but is not limited to, the following steps:

[0244] The first score is determined based on the motion gain data;

[0245] The second score is determined based on the number of actions performed.

[0246] Based on the first score and the second score, the prediction score of the molecular prediction result is determined.

[0247] Specifically, the process of determining the first score based on action gain data and the process of determining the second score based on the number of actions are similar to the process of determining the predicted score based on action gain data in step 1010 above. To save space, they will not be described in detail here.

[0248] Further, firstly, a first weight and a second weight are determined, where the sum of the first weight and the second weight is 1. The first weight is used to indicate the importance of the action gain data in identifying candidate biomolecules. The second weight is used to indicate the importance of the number of actions in identifying candidate biomolecules. Next, the product of the first weight and the first score is added to the product of the second weight and the second score to obtain the prediction score of the molecular prediction result.

[0249] The advantage of this embodiment is that it takes into account the impact of the number of actions during prediction optimization and the action gain data corresponding to the predicted biomolecules generated after each action change on molecular optimization. By combining the factors of the number of actions and action gain data to calculate the prediction score of each predicted biomolecule, the accuracy of the prediction score calculation can be improved.

[0250] The following is combined Figures 13A-13C An example is given to describe the specific process of determining the prediction score of molecular prediction results in a round.

[0251] like Figure 13A As shown, for the second biomolecule S0 with the molecular sequence [ASPMSL], a molecular optimization model was used to optimize and predict the second biomolecule S0, resulting in two predicted biomolecules: predicted biomolecule 1 with the molecular sequence [ASFMSL] and predicted biomolecule 2 with the molecular sequence [ASPSSL]. Predicted biomolecule 1 is formed by changing the amino acid at the third sequence position of the second biomolecule from P to F, and predicted biomolecule 2 is formed by changing the amino acid at the fourth sequence position of the second biomolecule from M to S. Next, for predicted biomolecule 2, the molecular optimization model was used to optimize and predict two new predicted biomolecules: predicted biomolecule 3 with the molecular sequence [AEPSSL] and predicted biomolecule 4 with the molecular sequence [ASPSSM]. Predicted biomolecule 3 is formed by changing the amino acid at the second sequence position of predicted biomolecule 2 from S to E, and predicted biomolecule 4 is formed by changing the amino acid at the sixth sequence position of predicted biomolecule 2 from L to M. Meanwhile, predicted biomolecule 1 was also optimized and predicted using the molecular optimization model, but no new molecular sequence was predicted. Furthermore, since the action gain data for both the predicted biomolecule 3 and the predicted biomolecule 4 are smaller than the action gain data for the predicted biomolecule 2, the optimization prediction for this round ends.

[0252] like Figure 13BThe diagram illustrates the optimized prediction of the second biomolecule (ASPMSL). Specifically, four predicted biomolecules were identified: Biomolecule 1 (1 action, action gain 76); Biomolecule 2 (1 action, action gain 88); Biomolecule 3 (2 actions, action gain 82); and Biomolecule 4 (2 actions, action gain 78).

[0253] like Figure 13C The diagram illustrates the prediction scores for the predicted biomolecules corresponding to the second biomolecule (ASPMSL). Specifically, when both the first and second weights are set to 0.5, the predicted biomolecule 1, with a first sub-score of 90 and a second sub-score of 76, has a prediction score of 83; the predicted biomolecule 2, with a first sub-score of 90 and a second sub-score of 88, has a prediction score of 89; the predicted biomolecule 3, with a first sub-score of 80 and a second sub-score of 82, has a prediction score of 81; and the predicted biomolecule 4, with a first sub-score of 80 and a second sub-score of 78, has a prediction score of 79. Based on this, predicted biomolecule 2 is selected as the candidate biomolecule for this round.

[0254] Step 330 will be described in detail below.

[0255] In step 330, reward predictions are performed on multiple candidate biomolecules based on a pre-trained prediction model to obtain the reward prediction results for each candidate biomolecule.

[0256] Please refer to Figure 14 In one embodiment, step 330 specifically includes, but is not limited to, the following steps 1410-1440:

[0257] Step 1410: Determine the first molecular sequence information of the first biomolecule and the second molecular sequence information of each of the multiple candidate biomolecules;

[0258] Step 1420: Extract the first sequence feature from the first molecular sequence information and extract the second sequence feature from the second molecular sequence information;

[0259] Step 1430: Based on the first sequence features and the second sequence features, a reward prediction is performed using a prediction model to obtain the molecular docking scores of each of the multiple candidate biomolecules with the first biomolecule.

[0260] Step 1440: Based on the molecular docking score, determine the reward prediction results for each of the multiple candidate biomolecules.

[0261] Steps 1410-1440 are described in detail below.

[0262] In step 1410, the first molecular sequence information is used to indicate the total number of molecular elements contained in the first biomolecule and the element type of the molecular elements at each sequence position. The second molecular sequence information is used to indicate the total number of molecular elements contained in the second biomolecule and the element type of the molecular elements at each sequence position.

[0263] For example, when the first biomolecule is a target protein and the second biomolecule is a polypeptide, the second molecular sequence information of the second biomolecule can characterize the total number of amino acids contained in the second biomolecule and the specific type of amino acids at each sequence position.

[0264] In this specific implementation, the first molecular sequence information of the first biomolecule is often recorded in a predetermined database. Therefore, with authorization, the first molecular sequence information of the first biomolecule can be found by searching the preset database using its name or serial number. These preset databases include, but are not limited to, databases providing protein sequence and functional information (UniProt), databases providing bioinformatics resources (National Center for Biotechnology Information, NCBI), or databases storing the three-dimensional structures of biological macromolecules (Protein Data Bank, PDB).

[0265] Furthermore, since the candidate biomolecules are predicted through molecular optimization models, their molecular structure and molecular characteristics can be analyzed to obtain their second molecular sequence information.

[0266] In step 1420, the first sequence feature is used to indicate the representation of the first molecular sequence information in the latent vector space. The second sequence feature is used to indicate the representation of the second molecular sequence information in the latent vector space.

[0267] In this specific implementation, the first molecular sequence information is mapped to a preset latent vector space to obtain a first sequence feature. Simultaneously, the second molecular sequence information is mapped to the preset latent vector space to obtain a second sequence feature.

[0268] In step 1430, the molecular docking score is used to assess the stability and affinity of molecular binding between biomolecules. The molecular docking score includes a combination of various interaction energies, such as Coulomb interaction energy, van der Waals interaction energy, and hydrogen bond interaction energy. Generally, a lower molecular docking score indicates a more stable binding between biomolecules (e.g., between a molecular ligand and a target protein) and higher affinity.

[0269] In this specific implementation, firstly, for each candidate biomolecule, the second sequence feature of the candidate biomolecule and the first sequence feature of the first biomolecule are concatenated to obtain a concatenated sequence feature. Next, the concatenated sequence feature is input into a prediction model for reward prediction to obtain the molecular docking score between the candidate biomolecule and the first biomolecule.

[0270] In this process, when inputting the spliced ​​sequence features into the prediction model for reward prediction, firstly, the convolutional layers of the prediction model are used to convolve the spliced ​​sequence features, capturing multi-scale sequence feature information of the first and second biomolecules to obtain molecular convolutional features. Next, self-attention is calculated on the molecular convolutional features to obtain the self-attention calculation result. Finally, reward prediction is performed based on the self-attention calculation result to obtain the molecular docking scores of each of the multiple candidate biomolecules with the first biomolecule.

[0271] Furthermore, since the three-dimensional structural information of the first and second biomolecules is crucial molecular information in affinity prediction, the molecular structural information of both biomolecules can be extracted for reward prediction. This information is then converted into latent space vectors and concatenated into molecular structure splicing features. These features, along with the splicing sequence features, are input into the prediction model for reward prediction. This allows the prediction model to jointly predict the molecular docking score between the candidate biomolecule and the first biomolecule based on both molecular sequence and structural features.

[0272] In step 1440, for each candidate biomolecule, the reciprocal of the molecular docking score can be directly used as the reward prediction result for that candidate biomolecule. The smaller the molecular docking score, the larger the reward prediction result.

[0273] Furthermore, to more accurately measure the reward of each candidate biomolecule based on its molecular docking score, a reward function can be defined. The docking molecule score is input into the preset reward function, and the output of the reward function is used as the reward prediction result for the candidate biomolecule. The preset reward function is a decreasing function with the molecular docking score as the independent variable and the reward prediction result of the candidate biomolecule as the dependent variable.

[0274] The advantage of this embodiment is that, considering the optimization of more target biomolecules with better binding affinity to the first biomolecule, the prediction model is set as a neural network model mainly based on predicting the binding affinity between molecules. During reward prediction, based on the molecular sequence information and three-dimensional structure information of the first biomolecule and candidate biomolecules, the molecular docking score of each candidate biomolecule with the first biomolecule is predicted, and the molecular docking score is converted into the reward prediction result of each candidate biomolecule, which can improve the accuracy of reward prediction.

[0275] In this embodiment of the disclosure, the prediction model includes a first prediction sub-model, a second prediction sub-model, and a third prediction sub-model.

[0276] The first predictive sub-model refers to the neural network model used to predict the binding affinity between the candidate biomolecule and the first biomolecule.

[0277] The second predictive sub-model refers to the neural network model used to predict the solubility of candidate biomolecules.

[0278] The third predictive sub-model refers to the neural network model used to predict the toxicity index (Toxicity) of candidate biomolecules.

[0279] It should be noted that the model structures of the first, second, and third prediction sub-models can be the same or different.

[0280] Please refer to Figure 15 In one embodiment, step 330 specifically includes, but is not limited to, the following steps 1510-1540:

[0281] Step 1510: For each candidate biomolecule, predict the binding affinity between the candidate biomolecule and the first biomolecule based on the first prediction sub-model, and determine the first reward result of the candidate biomolecule based on the binding affinity.

[0282] Step 1520: Predict the solubility of candidate biomolecules based on the second prediction sub-model, and determine the second reward result of candidate biomolecules based on the solubility;

[0283] Step 1530: Predict the toxicity index of candidate biomolecules based on the third prediction sub-model, and determine the third reward result of candidate biomolecules based on the toxicity index.

[0284] Step 1540: Based on the first reward result, the second reward result, and the third reward result, determine the reward prediction result of the candidate biomolecule.

[0285] Steps 1510-1540 are described in detail below.

[0286] In step 1510, binding affinity is used to indicate the stability and magnitude of the interaction between the candidate biomolecule and the first biomolecule. A first reward result is used to reward the candidate biomolecule based on binding affinity.

[0287] In the specific implementation of this embodiment, the process of step 1510 is similar to that of steps 1410-1440 described above. To save space, it will not be described again.

[0288] Furthermore, when determining the first reward result for candidate biomolecules based on binding affinity, a preset function can be invoked. The binding affinity is input into the preset function, and the output of the preset function is used as the first reward result. The preset function is an increasing function with binding affinity as the independent variable and the first reward result as the dependent variable.

[0289] In step 1520, solubility is used to indicate the water solubility of candidate biomolecules. The second reward result is used to assign a reward to the candidate biomolecules based on solubility.

[0290] In this specific implementation, the second prediction sub-model can be a neural network model based on multiple linear regression, robust regression, partial least squares regression, ridge regression, etc. Specifically, when the second prediction sub-model is a neural network model based on multiple linear regression, firstly, the second prediction sub-model extracts features from the candidate biomolecules, captures the feature vectors of the candidate biomolecules related to solubility, and then predicts the solubility of the candidate biomolecules based on the linear equation in the second prediction sub-model and the extracted feature vectors.

[0291] Furthermore, the specific process for determining the second reward result of candidate biomolecules based on solubility is similar to the specific process for determining the first reward result of candidate biomolecules based on binding affinity, as described above. To save space, it will not be elaborated further.

[0292] In step 1530, the toxicity index is used to indicate the toxicity level of the candidate biomolecule. The third reward result is used to reward the candidate biomolecule based on the toxicity index.

[0293] In this specific implementation, the third prediction sub-model can be a neural network model composed of a graph neural network (GNN), a convolutional neural network (CNN), and a bidirectional long short-term memory network (BiLSTM). Specifically, when using the third prediction sub-model to predict the toxicity of candidate biomolecules, firstly, the graph neural network of the third prediction sub-model is used to extract molecular-level features from the molecular graph of the candidate biomolecule. Next, the convolutional neural network and the bidirectional long short-term memory network are used to extract the evolutionary features of the candidate biomolecule from the position-specific score matrix. Further, the molecular-level features and the evolutionary features are concatenated to obtain molecular concatenated features. Then, self-attention calculation is performed on the molecular concatenated features to obtain feature attention results, and toxicity prediction is performed based on the feature attention results to obtain the toxicity index of the candidate biomolecule.

[0294] Furthermore, the specific process for determining the third reward result of candidate biomolecules based on toxicity indicators is similar to the specific process for determining the first reward result of candidate biomolecules based on binding affinity, as described above. To save space, it will not be elaborated further.

[0295] In step 1540, the first reward result, the second reward result, and the third reward result are averaged to obtain the reward prediction result of the candidate biomolecule.

[0296] like Figure 16 The diagram illustrates a simplified method for predicting rewards for each candidate biomolecule using a prediction model. Specifically, the molecular sequence information of the second biomolecule S0, candidate biomolecule S1, candidate biomolecule S2, ..., candidate biomolecule Sn is input into the prediction model to obtain the reward prediction result r0 for the second biomolecule S0, the reward prediction result r1 for the candidate biomolecule S1, the reward prediction result r2 for the candidate biomolecule S2, ..., and the reward prediction result rn for the candidate biomolecule Sn.

[0297] The advantage of this embodiment is that it utilizes a predictive model to predict molecular properties across multiple dimensions, such as binding affinity, toxicity indices, and solubility, thereby improving the diversity of predictions. Simultaneously, by determining the reward outcomes for candidate biomolecules across multiple dimensions based on binding affinity, toxicity indices, and solubility, and by comprehensively considering the reward outcomes from these multiple dimensions to jointly determine the reward prediction result for the candidate biomolecule, the accuracy of reward prediction can be improved.

[0298] Step 340 will be described in detail below.

[0299] In step 340, the target biomolecule is identified from multiple candidate biomolecules based on the reward prediction results.

[0300] Please refer to Figure 17In one embodiment, step 340 specifically includes, but is not limited to, the following steps 1710-1720:

[0301] Step 1710: Based on the magnitude of the reward prediction results, sort multiple candidate biomolecules in descending order to obtain molecular sequences;

[0302] Step 1720: Identify the first predetermined number of candidate biomolecules in the molecular sequence as the target biomolecule.

[0303] Steps 1710-1720 are described in detail below.

[0304] In step 1710, the molecular sequence is used to indicate the result of sorting multiple candidate biomolecules according to the reward prediction results.

[0305] In this specific implementation, based on the reward prediction results of each candidate biomolecule, the candidate biomolecules are arranged in descending order of their reward prediction results to obtain a molecular sequence. The candidate biomolecules ranked earlier in the molecular sequence have higher reward prediction results, while those ranked later have lower reward prediction results.

[0306] In step 1720, the predetermined number is used to indicate the total number of target biomolecules to be identified.

[0307] In this specific implementation, a higher reward prediction result for a candidate biomolecule indicates a better binding affinity between the candidate biomolecule and the first biomolecule, making it more suitable as the target biomolecule. Based on this, a predetermined number of candidate biomolecules in the molecular sequence are selected as the target biomolecules according to a pre-set number.

[0308] like Figure 18The diagram illustrates the process of screening for a target biomolecule from multiple candidate biomolecules. Specifically, when the first biomolecule is a target protein, there are eight candidate biomolecules (candidate binding peptides) corresponding to that target protein: candidate biomolecule 1 (predicted reward result 63), candidate biomolecule 2 (predicted reward result 74), candidate biomolecule 3 (predicted reward result 78), candidate biomolecule 4 (predicted reward result 82), candidate biomolecule 5 (predicted reward result 77), candidate biomolecule 6 (predicted reward result 86), candidate biomolecule 7 (predicted reward result 84), and candidate biomolecule 8 (predicted reward result 90). Based on this, these eight candidate biomolecules are arranged in descending order according to their prediction results, forming the following sequence (molecular sequence): [candidate biomolecule 8, candidate biomolecule 6, candidate biomolecule 7, candidate biomolecule 4, candidate biomolecule 3, candidate biomolecule 5, candidate biomolecule 2, candidate biomolecule 1]. When the predetermined number is three, the target biomolecules are candidate biomolecules 8, 6, and 7.

[0309] The advantage of this embodiment is that, for multiple candidate biomolecules, reward prediction is performed using a prediction model, and the molecular performance of each candidate biomolecule is indicated based on the reward prediction results. At the same time, the multiple candidate molecules are arranged in descending order according to the magnitude of the reward prediction results, which can more clearly and intuitively reflect the differences in the molecular performance of each candidate biomolecule, thereby improving the rationality and accuracy of screening target biomolecules with better affinity binding to the first biomolecule from multiple candidate biomolecules.

[0310] The specific process of optimizing the prediction model according to an embodiment of this disclosure will be described in detail below.

[0311] Since the accuracy of prediction models is not absolute, and they do not provide absolutely accurate predictions for every target biomolecule, they may occasionally produce incorrect reward predictions for some target biomolecules. Therefore, relying entirely on the reward predictions output by the model often leads to errors in recording the binding affinity between some target biomolecules and the first biomolecule, hindering the optimization of the prediction model by constructing new sample data based on the target biomolecules. To address this, this disclosure provides an experimentally validated model optimization scheme that can verify the correctness of the reward predictions for target biomolecules, thereby improving the accuracy of model optimization.

[0312] Please refer to Figure 19 In one embodiment, the specific process of optimizing the prediction model includes, but is not limited to, the following steps 1910-1920:

[0313] Step 1910: Perform molecular binding verification on the target biomolecule and the first biomolecule based on the preset verification rules to obtain the verification results;

[0314] Step 1920: In response to the discrepancy between the verification results and the reward prediction results of the target biomolecule, update the parameters of the prediction model based on the verification results, the target biomolecule, and the first biomolecule.

[0315] Steps 1910-1920 are described in detail below.

[0316] In step 1910, the preset verification rules refer to various experimental verification methods applicable to biomolecules, such as wet experimental verification, structural evaluation, binding energy evaluation, or molecular dynamics simulation. The verification results are used to indicate the binding affinity between the target biomolecule and the first biomolecule obtained according to the preset verification rules.

[0317] In this specific implementation, firstly, the verification operation steps and analysis strategy of the preset verification rules are determined. Next, according to the verification operation steps, molecular binding verification operations are performed on the target biomolecule and the first biomolecule in the laboratory to obtain experimental data. Further, the experimental data is analyzed according to the analysis strategy to obtain verification results, which are then used to verify the accuracy of the prediction model's reward prediction results for the target biomolecule.

[0318] In step 1920, firstly, for the target biomolecule, the verification result is compared with the reward prediction result of the target biomolecule. Next, if the verification result is inconsistent with the reward prediction result of the target biomolecule, it indicates that the prediction model's prediction for that target biomolecule is incorrect. The verification result is then used to replace the reward prediction result of the target biomolecule, serving as a reference label for the target biomolecule. The target biomolecule, the first biomolecule, and the verification result are then input into the prediction model as a new sample molecule pair to update the model's parameters. The model parameters are continuously adjusted to ensure that the prediction model's prediction result for this new sample molecule pair is as close as possible to the verification result.

[0319] The advantage of this embodiment is that it introduces preset verification rules for molecular binding verification in the laboratory for both the target biomolecule and the first biomolecule. This allows for the verification of the correctness of the prediction model's reward prediction results for the target biomolecule through real experimental validation, enabling timely correction of the prediction results between the target biomolecule and the first biomolecule. Furthermore, by treating the target biomolecule, the first biomolecule, and the verification result as a new sample molecule pair, it achieves continuous optimization of the prediction model by constructing new sample molecule pairs during model application. This improves the accuracy of model optimization while enhancing the accuracy of the new sample molecule pairs, thereby continuously improving the predictive performance of the prediction model.

[0320] The following is a detailed description of the specific process of training a prediction model according to an embodiment of this disclosure.

[0321] Please refer to Figure 20 In one embodiment, the specific process of training the prediction model includes, but is not limited to, the following steps 2010-2050:

[0322] Step 2010: Obtain multiple sample molecule pairs;

[0323] Step 2020: Extract the first sample sequence features from the molecular sequence information of the first sample biomolecule and extract the second sample sequence features from the molecular sequence information of the second sample biomolecule.

[0324] Step 2030: Fuse the first sample sequence features and the second sample sequence features into a sample sequence fusion feature;

[0325] Step 2040: Based on the sample sequence fusion features, predict the binding affinity of sample molecule pairs using the original model to obtain the predicted binding data of the first sample biomolecule and the second sample biomolecule in the sample molecule pair.

[0326] Step 2050: Based on the reference combined data and the prediction combined data, train the original model to obtain the prediction model.

[0327] Steps 2010-2050 are described in detail below.

[0328] In step 2010, the sample molecule pair includes a first sample biomolecule, a second sample biomolecule, and reference binding data between the first sample biomolecule and the second sample biomolecule.

[0329] The first sample biomolecule refers to the biomolecule required during the model training phase of the prediction model. The first sample biomolecule is of the same type as the first biomolecule.

[0330] The second sample biomolecule refers to the biomolecule selected during the model training phase of the prediction model that can bind to the first sample biomolecule. The second sample biomolecule is of the same type as the second biomolecule.

[0331] Reference binding data are used to indicate the true binding affinity of the first sample biomolecule and the second sample biomolecule.

[0332] In this specific implementation, since the training objective of the prediction model is to enable it to predict the binding affinity between biomolecules, the sample data used for model training needs to be binding affinity data of biomolecules. Based on this, with authorization, affinity data can be extracted as sample molecule pairs from databases specifically targeting the binding affinity of target proteins and peptides, the binding affinity of membrane proteins and protein complexes, or the binding affinity of proteins and proteins. Among these, the binding affinity databases permitted to be used in this embodiment include, but are not limited to, the MPAD database, the PPB-Affinity database, etc.

[0333] In step 2020, the first sample sequence feature is used to indicate the representation of the molecular sequence information of the first sample biomolecule in the latent vector space. The second sample sequence feature is used to indicate the representation of the molecular sequence information of the second sample biomolecule in the latent vector space.

[0334] In the specific implementation of this embodiment, the process of step 2020 is similar to that of steps 1410-1420 described above. To save space, it will not be described again.

[0335] In step 2030, the sample sequence fusion features are used to comprehensively characterize the molecular sequence information of the first sample biomolecule and the molecular sequence information of the second sample biomolecule.

[0336] In the specific implementation of this embodiment, for each sample molecule pair, the first sample sequence features and the second sample sequence features are spliced ​​together to achieve information fusion of the molecular sequence information of the first sample biomolecule and the molecular sequence information of the second sample biomolecule, thereby obtaining the spliced ​​sequence features.

[0337] In step 2040, the original model refers to an untrained neural network model used to predict the binding affinity between biomolecules. The original model can be a neural network model based on fully connected neural networks, convolutional neural networks, Transformer networks, etc.

[0338] The predicted binding data is used to indicate the binding affinity of the first and second sample biomolecules predicted by the original model.

[0339] In the specific implementation of this embodiment, the process of step 2040 is similar to that of steps 1430-1440 described above. To save space, it will not be described again.

[0340] In step 2050, the specific process of training the original model based on the reference combined data and the prediction combined data will be described in detail below, and will not be repeated here.

[0341] The advantage of this embodiment is that it takes into account the differences between the predicted binding data and the reference binding data generated for each sample molecule pair in each iteration of the original model training. This approach can train the original model by minimizing the differences between the predicted binding data and the reference binding data generated for each sample molecule pair, which is beneficial to improving the model training effect and thus improving the affinity prediction accuracy of the trained prediction model.

[0342] Please refer to Figure 21 In one embodiment, the specific process of training the prediction model includes, but is not limited to, the following steps 2110-2130:

[0343] Step 2110: For each sample pair, determine the sub-loss function based on the reference binding data and the predicted binding data;

[0344] Step 2120: Determine the total loss function based on the sub-loss functions of multiple sample pairs;

[0345] Step 2130: Based on the total loss function, train the original model to obtain the prediction model.

[0346] Steps 2110-2130 are described in detail below.

[0347] In step 2110, the sub-loss function is used to indicate the degree of difference between the original model's prediction of the binding affinity between the first and second sample biomolecules for a single sample molecule pair and the actual binding affinity.

[0348] In this specific implementation, since both the reference binding data and the predicted binding data refer to the binding affinity between the first sample biomolecule and the second sample biomolecule, and binding affinity is often described numerically as the strength of the binding between the sample biomolecule and the second sample biomolecule, for each sample molecule pair, the reference binding data and the predicted binding data can be subtracted to obtain the binding affinity difference. Then, the binding affinity difference is squared to obtain the sub-loss function for the sample molecule pair.

[0349] In step 2120, the total loss function is used to indicate the overall difference between the original model's prediction of the binding affinity between the first and second sample biomolecules for multiple sample molecule pairs and the actual binding affinity.

[0350] In the specific implementation of this embodiment, firstly, the total number of sample pairs is determined. Next, based on the total number of sample pairs, the sub-loss functions of multiple sample pairs are averaged, and the total loss function is obtained by dividing the sum of the sub-loss functions of multiple sample pairs by the total number.

[0351] In step 2130, with the goal of minimizing the total loss function, the model parameters of the original model are adjusted, and steps 2010-2050 above are repeated to achieve iterative training of the original model. The model parameters that minimize the total loss function are taken as the final model parameters, and the original model with the final model parameters is taken as the trained prediction model.

[0352] The advantage of this embodiment is that it takes into account the differences between the predicted binding data and the reference binding data generated by the original model for each sample molecule pair in each iteration of training. Based on supervised learning, the total loss function is jointly determined according to the differences between the predicted binding data and the reference binding data for each sample molecule pair. This approach can train the original model by minimizing the differences between the predicted binding data and the reference binding data for each sample molecule pair, which is beneficial to improving the model training effect and thus improving the affinity prediction accuracy of the trained prediction model.

[0353] The following describes the implementation details of the molecular optimization method according to embodiments of this disclosure.

[0354] The following reference Figure 22 The following provides a detailed and exemplary description of the implementation details of the molecular optimization method according to embodiments of the present disclosure.

[0355] like Figure 22As shown, firstly, affinity experiment data can be extracted from the affinity experiment database as sample molecule pairs. These sample molecule pairs are then used to fine-tune the original model to obtain a predictive model, a process similar to steps 2010-2050 above. Next, the target protein is used as the first biomolecule, and the initial peptide as the second biomolecule. The agent in the peptide optimization generator and the predictive model are used to search for and optimize candidate peptides, resulting in multiple candidate peptides, a process similar to step 320 above. After obtaining the candidate peptides, the predictive model is used again to perform reward scoring and screening, selecting the candidate peptides with higher reward prediction results as the target peptides, a process similar to steps 330-340 above. Further, experimental validation is performed on the target peptides. This validation can include wet experiments, structural evaluation, etc. The results of the experimental validation can be used to further optimize the predictive model, a process similar to steps 1910-1920 above. Finally, the optimized predictive model is used to perform a new round of optimization on the optimal peptide sequence obtained in the aforementioned process, thereby continuously improving the performance and quality of the candidate peptides.

[0356] The experimental verification of the molecular optimization method according to the embodiments of this disclosure is described below with examples.

[0357] The following reference Figures 23A-23C The effectiveness of the molecular optimization method of the embodiments of this disclosure will be explained in detail by way of example.

[0358] The biooptimization method based on embodiments of this disclosure performs peptide generation and optimization experiments on four biologically significant target proteins. The four target proteins are glucagon-like peptide-1 receptor (GLP-1R), MCL1 protein, EDB protein, and DH270.6 protein. Specifically, firstly, for each target protein, a random sequence is initialized as the starting point for molecular optimization. The target protein is considered the first biomolecule, and the random sequence is considered the second biomolecule. The random sequence is... Figure 23A The results are indicated by green circles. Next, each initial random sequence was iteratively optimized 1000 times, with each iteration aiming to improve the binding affinity between the random sequence and the target protein. This process generated 1000 candidate sequences, providing sufficient data to evaluate the evolution of the random sequences. To improve the reliability of the experimental results, a known binding native peptide was selected as a control group for each target protein, serving as the experimental baseline to compare the differences and similarities between the optimized sequence and the binding native peptide.

[0359] like Figure 23AAs shown, based on the positional differences between the optimized sequences corresponding to the four target proteins and the bound natural peptides, in the initial stage of the experiment, the optimized sequences generated by the molecular optimization method of this disclosure were significantly different from the bound natural peptides. However, as the optimization process progressed, it was observed that the optimized sequences gradually moved closer to the distribution region of the bound natural peptides. This indicates that the molecular optimization method of this disclosure can effectively guide random sequences to mutate in the correct direction.

[0360] like Figure 23B The figure illustrates the change in similarity between the optimized sequences and the natural sequences generated during the optimization process. As the number of iterations increases, the similarity between the optimized sequences corresponding to the four target proteins and the sequences binding to the natural peptides gradually increases, eventually stabilizing at approximately 50%. This result not only demonstrates that the molecular optimization method of this disclosure can optimize sequences to a spatial level close to that of natural peptides, but also generates polypeptide sequences with a certain degree of novelty.

[0361] like Figure 23C The figure illustrates the evaluation of the evolution ratio of peptide sequences after 1000 iterations, where the evolution ratio refers to the proportion of amino acid mutations in the peptide sequence relative to the random sequence. As shown in the figure, almost all target protein peptide sequences underwent more than 80% evolution. This finding indicates that the molecular optimization method of this disclosure can not only perform subtle single-point mutations but also significantly evolve the entire sequence, thereby forming novel binding peptide sequences.

[0362] The following reference Figures 24A-24E The following detailed and exemplary description illustrates the changing trends of the physicochemical properties of multiple candidate biomolecules generated during the sequence optimization process corresponding to the molecular optimization method of this disclosure embodiment.

[0363] like Figures 24A-24E As shown, this disclosure optimizes five physicochemical properties of peptides targeting the aforementioned four target proteins (glucagon-like peptide-1 receptor, MCL1 protein, EDB protein, and DH270.6 protein). These five physicochemical properties include aromaticity, charge, isoelectric point, global hydrophobicity, and global hydrophobic moment. These physicochemical properties often directly affect the interaction between the peptide and its target, as well as its distribution and metabolism in vivo. Figures 24A-24E In the diagram, the red dashed line represents the physicochemical properties of the bound natural peptide. Using these properties as a benchmark, it is possible to intuitively compare the differences in various physicochemical properties between the optimized peptides and the bound natural peptides at different iteration stages. Figures 24A-24EAs can be seen, most peptides, during the optimization process, exhibit physicochemical properties that can generate or approximate the binding properties of native peptides. This observation indicates that the optimization strategies of the present disclosure not only improve the binding affinity of peptides but also maintain or enhance the key physicochemical properties required for their biological functions. Furthermore, the molecular optimization methods of the present disclosure can generate peptide sequences with a variety of physicochemical properties, demonstrating the flexibility and diversity of the optimization strategies. This diversity is important for adapting to different biological needs and therapeutic goals. For example, by adjusting the global hydrophobicity of a peptide, its permeability in cell membranes can be optimized, thereby improving drug bioavailability and efficacy.

[0364] The following reference Figure 25 The effectiveness of iterative optimization of the molecular optimization method in the embodiments of this disclosure will be explained in detail by way of example.

[0365] like Figure 25 As shown in the embodiments of this disclosure, a GLP-1R target protein was selected as the first biomolecule, and a peptide with a relatively low initial affinity to the first biomolecule was selected as the second biomolecule. A series of optimization experiments were conducted on this peptide with relatively low initial affinity, where the initial affinity was approximately 10⁻⁴. Specifically, in this experiment, a first round of iterative optimization was performed on this low-affinity peptide, with a total of 1000 iterations. Each iteration aimed to improve the binding affinity of the peptide to the GLP-1R target protein. According to... Figure 25 The left side of the diagram illustrates that a significant upward trend in affinity can be observed throughout the optimization process. Ultimately, the affinity successfully increased to approximately 10^-5, a considerable improvement. Furthermore, based on the first round of optimization, the optimized sequence was further evolved to achieve even higher affinity. Figure 25 The right side of the diagram illustrates that after the second round of optimization, the affinity of the peptide sequence was further improved, reaching approximately 10^-6. This result demonstrates that the optimization strategy of this disclosure not only significantly improves peptide affinity but also achieves continuous improvement. Through two carefully designed optimization processes, the peptide affinity was successfully increased from 10^-4 to 10^-6, a significant improvement of approximately 100 times. This achievement not only proves the effectiveness of the molecular optimization method of this disclosure but also demonstrates the powerful potential of deep reinforcement learning in the field of peptide optimization.

[0366] The apparatus and device according to embodiments of this disclosure will now be described.

[0367] It is understood that although the steps in the above flowcharts are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated in this embodiment, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the above flowcharts may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0368] It should be noted that in various specific embodiments of this application, when processing is required based on data related to the characteristics of the target object, such as target object attribute information or a set of attribute information, the permission or consent of the target object will be obtained first. Furthermore, the collection, use, and processing of this data will comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require obtaining target object attribute information, separate permission or consent from the target object will be obtained through pop-ups or redirection to a confirmation page. Only after obtaining the target object's separate permission or consent will the necessary target object-related data for the normal operation of the embodiments of this application be obtained.

[0369] Figure 26 A schematic diagram of the structure of a molecular optimization device 2600 provided in an embodiment of this disclosure. The molecular optimization device 2600 includes:

[0370] The first determining unit 2610 is used to determine the first biomolecule and the second biomolecule to be optimized;

[0371] The optimization unit 2620 is used to optimize and predict the second biomolecule based on the first biomolecule using a preset molecular optimization model, and to determine the candidate biomolecules predicted in each round, wherein the candidate biomolecules in each round are predicted based on the candidate biomolecules in the previous round relative to the round.

[0372] The prediction unit 2630 is used to predict the reward of multiple candidate biomolecules based on a pre-trained prediction model, and obtain the reward prediction results of each candidate biomolecule.

[0373] The second determining unit 2640 is used to determine the target biomolecule from multiple candidate biomolecules based on the reward prediction result, wherein the binding affinity between the target biomolecule and the first biomolecule is greater than the binding affinity between the second biomolecule and the first biomolecule.

[0374] Optionally, the optimization unit 2620 includes:

[0375] An initialization module (not shown) is used to initialize the current round number to 1 and the biomolecule to be examined to the second biomolecule.

[0376] The first determining module (not shown) is used to perform optimization prediction on the biomolecule under investigation based on the first biomolecule in the current round, and determine multiple molecular prediction results and the action gain data of each of the multiple molecular prediction results.

[0377] The second determination module (not shown) is used to identify candidate biomolecules from multiple molecular prediction results based on action gain data.

[0378] The loop module (not shown) is used to increment the round number by 1, update the biomolecule to be examined to the candidate biomolecule, and return to the steps in the current round, based on the first biomolecule, to optimize and predict the biomolecule to be examined through the molecular optimization model, determine multiple molecular prediction results, and the action gain data of each of the multiple molecular prediction results, until the round number is the target number, and multiple candidate biomolecules are obtained.

[0379] Optionally, the first determining module (not shown) includes:

[0380] An extraction submodule (not shown) is used to extract the molecular feature information of the first biomolecule in the current round.

[0381] The prediction submodule (not shown) is used to optimize and predict the biomolecule under investigation based on molecular feature information through a molecular optimization model, obtain the predicted biomolecule, and determine the action gain data of the predicted biomolecule based on the difference between the predicted biomolecule and the second biomolecule.

[0382] The loop submodule (not shown) is used to update the biomolecule under investigation with the predicted biomolecule and return the steps of optimizing the prediction of the biomolecule under investigation based on molecular feature information and through the molecular optimization model to obtain the predicted biomolecule until the predicted biomolecule and the action gain data meet the first condition, the predicted biomolecule is determined as the molecular prediction result, and the action gain data of the predicted biomolecule is determined as the action gain data of the molecular prediction result.

[0383] Optionally, the prediction submodule (not shown) is used for:

[0384] Based on the difference between the predicted biomolecule and the second biomolecule, the action space matrix for the transformation from the second biomolecule to the predicted biomolecule is determined.

[0385] Based on the action space matrix and molecular feature information, gain prediction is performed through a molecular optimization model to obtain action gain data.

[0386] Optionally, the loop submodule (not shown) is used for:

[0387] Determine the action gain data of the biomolecule under investigation corresponding to the predicted biomolecule;

[0388] In response to the fact that the action gain data of the predicted biomolecules are all less than the action gain data of the biomolecule under investigation corresponding to the predicted biomolecules, it is determined that the predicted biomolecules and the action gain data meet the first condition.

[0389] Optionally, the loop submodule (not shown) is used for:

[0390] A control biomolecule is determined based on the predicted biomolecule and the second biomolecule generated prior to the prediction biomolecule.

[0391] In response to the prediction biomolecule being identical to the control biomolecule, the prediction biomolecule and action gain data are determined to meet the first condition.

[0392] Optionally, the second determining module (not shown) includes:

[0393] The first determining submodule (not shown) is used to determine the prediction score of each molecule prediction result based on the action gain data.

[0394] The second determination submodule (not shown) is used to identify candidate biomolecules from multiple molecular prediction results based on prediction scores.

[0395] Optionally, the first determining submodule (not shown) is used for:

[0396] For each molecule prediction result, determine the number of actions required to transform the biomolecule under investigation into the molecule prediction result;

[0397] The prediction score of the molecular prediction result is determined based on action gain data and the number of actions.

[0398] Optionally, based on action gain data and the number of actions, a prediction score for the molecular prediction result is determined, including:

[0399] The first score is determined based on the motion gain data;

[0400] The second score is determined based on the number of actions performed.

[0401] Based on the first score and the second score, the prediction score of the molecular prediction result is determined.

[0402] Optionally, the prediction unit 2630 is used for:

[0403] Determine the first molecular sequence information of the first biomolecule and the second molecular sequence information of each of the multiple candidate biomolecules;

[0404] Extract first sequence features from first molecular sequence information, and extract second sequence features from second molecular sequence information;

[0405] Based on the first sequence features and the second sequence features, a prediction model is used to predict rewards and obtain the molecular docking scores of multiple candidate biomolecules with the first biomolecule.

[0406] Based on molecular docking scores, the reward prediction results for each of the multiple candidate biomolecules are determined.

[0407] Optionally, the prediction model includes a first prediction sub-model, a second prediction sub-model, and a third prediction sub-model;

[0408] Prediction unit 2630 is used for:

[0409] For each candidate biomolecule, the binding affinity between the candidate biomolecule and the first biomolecule is predicted based on the first prediction sub-model, and the first reward result of the candidate biomolecule is determined based on the binding affinity.

[0410] The solubility of candidate biomolecules is predicted based on the second predictor sub-model, and the second reward result of the candidate biomolecules is determined based on the solubility.

[0411] The toxicity index of candidate biomolecules is predicted based on the third predictor sub-model, and the third reward result of candidate biomolecules is determined based on the toxicity index.

[0412] Based on the first reward result, the second reward result, and the third reward result, the reward prediction result of the candidate biomolecule is determined.

[0413] Optionally, the second biomolecule to be optimized is obtained through the following methods:

[0414] Identify multiple reference biomolecules whose binding affinity to the first biomolecule is greater than a preset threshold;

[0415] Based on the molecular physicochemical information of multiple reference biomolecules, a second biomolecule is identified among the multiple reference biomolecules.

[0416] Optionally, the second determining unit 2640 is used for:

[0417] Based on the magnitude of the reward prediction results, multiple candidate biomolecules are sorted in descending order to obtain molecular sequences;

[0418] The first predetermined number of candidate biomolecules in the molecular sequence are identified as the target biomolecule.

[0419] Optionally, the molecular optimization device 2600 further includes an update unit (not shown), which is used for:

[0420] Based on preset verification rules, molecular binding verification of the target biomolecule and the first biomolecule is performed to obtain verification results;

[0421] In response to the discrepancy between the validation results and the reward prediction results of the target biomolecule, the prediction model parameters are updated based on the validation results, the target biomolecule, and the first biomolecule.

[0422] Optionally, the molecular optimization device 2600 further includes a training unit (not shown), which is used for:

[0423] Multiple sample molecule pairs are obtained, wherein each sample molecule pair includes a first sample biomolecule, a second sample biomolecule, and reference binding data of the first sample biomolecule and the second sample biomolecule;

[0424] Extract the first sample sequence features from the molecular sequence information of the first sample biomolecule, and extract the second sample sequence features from the molecular sequence information of the second sample biomolecule;

[0425] The features of the first sample sequence and the features of the second sample sequence are fused into a sample sequence fusion feature.

[0426] Based on the sample sequence fusion features, the binding affinity of sample molecule pairs is predicted using the original model, and the predicted binding data of the first sample biomolecule and the second sample biomolecule in the sample molecule pair are obtained.

[0427] The original model is trained based on the reference combined data and the prediction combined data to obtain the prediction model.

[0428] Optionally, based on the reference combined data and the prediction combined data, the original model is trained to obtain the prediction model, including:

[0429] For each sample pair, a sub-loss function is determined based on reference binding data and predicted binding data;

[0430] The total loss function is determined based on the sub-loss functions of multiple sample pairs;

[0431] The original model is trained based on the total loss function to obtain the prediction model.

[0432] Reference Figure 27 , Figure 27To implement the structural block diagram of the terminal portion of the molecular optimization method according to embodiments of this disclosure, the terminal includes: a radio frequency (RF) circuit 2710, a memory 2715, an input unit 2730, a display unit 2740, a sensor 2750, an audio circuit 2760, a wireless fidelity (WiFi) module 2770, a processor 2780, and a power supply 2790, etc. Those skilled in the art will understand that... Figure 27 The terminal structure shown does not constitute a limitation on mobile phones or computers and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0433] The RF circuit 2710 can be used to receive and transmit signals during information transmission or calls. In particular, it receives downlink information from the base station and processes it with the processor 2780; in addition, it transmits uplink data to the base station.

[0434] The memory 2715 can be used to store software programs and modules, and the processor 2780 executes various functional applications and data processing of the target terminal by running the software programs and modules stored in the memory 2715.

[0435] The input unit 2730 can be used to receive input numeric or character information, and to generate key signal inputs related to the settings and function control of the target terminal. Specifically, the input unit 2730 may include a touch panel 2731 and other input devices 2732.

[0436] Display unit 2740 can be used to display input or provided information, as well as various menus of the target terminal. Display unit 2740 may include display panel 2741.

[0437] Audio circuitry 2760, speaker 2761, and microphone 2762 provide an audio interface.

[0438] In this embodiment, the processor 2780 included in the terminal can execute the molecular optimization method of the previous embodiment.

[0439] The terminals disclosed in this embodiment include, but are not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, and aircraft. The embodiments of this invention can be applied to various scenarios, including but not limited to big data, biomedicine, data storage, and information technology.

[0440] Figure 28This is a partial structural block diagram of a server for implementing the molecular optimization method of this disclosure. The server can vary considerably depending on its configuration or performance, and may include one or more Central Processing Units (CPUs) 2822 (e.g., one or more processors) and a memory 2832, and one or more storage media 2830 (e.g., one or more mass storage devices) for storing application programs 2842 or data 2844. The memory 2832 and storage media 2830 may be temporary or persistent storage. The program stored in the storage media 2830 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the server. Furthermore, the CPU 2822 may be configured to communicate with the storage media 2830 and execute the series of instruction operations in the storage media 2830 on the server.

[0441] The server may also include one or more power supplies 2826, one or more wired or wireless network interfaces 2850, one or more input / output interfaces 2858, and / or one or more operating systems 2841, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0442] The central processing unit 2822 in the server can be used to execute the molecular optimization method of the embodiments of this disclosure.

[0443] This disclosure also provides a computer-readable storage medium for storing program code for executing the molecular optimization methods of the foregoing embodiments.

[0444] This disclosure also provides a computer program product comprising a computer program. A processor of a computer device reads and executes the computer program, causing the computer device to perform the aforementioned transaction on-chaining.

[0445] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in this disclosure and the foregoing drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “including,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatuses.

[0446] It should be understood that in this disclosure, "at least one item" means one or more, and "more than one" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0447] It should be understood that in the description of the embodiments disclosed herein, "multiple" means two or more, "greater than", "less than", "exceeding" etc. are understood to exclude the number itself, and "above", "below", "within" etc. are understood to include the number itself.

[0448] In the several embodiments provided in this disclosure, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0449] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0450] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0451] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0452] It should also be understood that the various implementation methods provided in this disclosure can be combined arbitrarily to achieve different technical effects.

[0453] The above is a detailed description of the embodiments of this disclosure. However, this disclosure is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this disclosure. All such equivalent modifications or substitutions are included within the scope defined by the claims of this disclosure.

Claims

1. A molecular optimization method, characterized in that, The method includes: Identify the first biomolecule and the second biomolecule to be optimized; Based on the first biomolecule, the second biomolecule is optimized and predicted using a preset molecular optimization model to determine candidate biomolecules predicted in each round, wherein the candidate biomolecules in each round are predicted based on the candidate biomolecules in the previous round relative to the round. Reward prediction is performed on multiple candidate biomolecules based on a pre-trained prediction model, and the reward prediction results for each of the multiple candidate biomolecules are obtained. Based on the reward prediction results, a target biomolecule is identified among the plurality of candidate biomolecules, wherein the binding affinity between the target biomolecule and the first biomolecule is greater than the binding affinity between the second biomolecule and the first biomolecule.

2. The molecular optimization method according to claim 1, characterized in that, The step of optimizing and predicting the second biomolecule based on the first biomolecule using a preset molecular optimization model to determine candidate biomolecules in each round of prediction includes: The current round number is initialized to 1, and the biomolecule to be examined is initialized to the second biomolecule. In the current round, based on the first biomolecule, the molecular optimization model is used to optimize and predict the biomolecule under investigation, and multiple molecular prediction results and the action gain data of each of the multiple molecular prediction results are determined. Based on the action gain data, the candidate biomolecules are determined from the multiple molecular prediction results; Increment the round number by 1, update the biomolecule to be examined to the candidate biomolecule, and return to the step of optimizing and predicting the biomolecule to be examined based on the first biomolecule in the current round using the molecular optimization model, determining multiple molecular prediction results and the action gain data of each of the multiple molecular prediction results, until the round number is the target number, thus obtaining multiple candidate biomolecules.

3. The molecular optimization method according to claim 2, characterized in that, In the current round, based on the first biomolecule, the molecular optimization model is used to optimize and predict the biomolecule under investigation, determining multiple molecular prediction results and the action gain data of each of the multiple molecular prediction results, including: In the current round, extract the molecular feature information of the first biomolecule; Based on the molecular feature information, the biomolecule under investigation is optimized and predicted using the molecular optimization model to obtain the predicted biomolecule. Based on the difference between the predicted biomolecule and the second biomolecule, the action gain data of the predicted biomolecule is determined. The predicted biomolecule is used to update the biomolecule to be examined, and the steps of optimizing and predicting the biomolecule to be examined based on the molecular feature information and the molecular optimization model are returned to obtain the predicted biomolecule, until the predicted biomolecule and the action gain data meet the first condition, the predicted biomolecule is determined as the molecular prediction result, and the action gain data of the predicted biomolecule is determined as the action gain data of the molecular prediction result.

4. The molecular optimization method according to claim 3, characterized in that, The step of determining the action gain data of the predicted biomolecule based on the difference between the predicted biomolecule and the second biomolecule includes: Based on the difference between the predicted biomolecule and the second biomolecule, the action space matrix for transforming from the second biomolecule to the predicted biomolecule is determined. Based on the action space matrix and the molecular feature information, the action gain data is obtained by performing gain prediction through the molecular optimization model.

5. The molecular optimization method according to claim 3, characterized in that, The determination that the predicted biomolecule and the action gain data meet the first condition includes: Determine the action gain data of the biomolecule under investigation corresponding to the predicted biomolecule; In response to the fact that the action gain data of the predicted biomolecule is less than the action gain data of the biomolecule under investigation corresponding to the predicted biomolecule, it is determined that the predicted biomolecule and the action gain data meet a first condition.

6. The molecular optimization method according to claim 3, characterized in that, The determination that the predicted biomolecule and the action gain data meet the first condition includes: A control biomolecule is determined based on the predicted biomolecule generated prior to the predicted biomolecule and the second biomolecule. In response to the prediction biomolecule being the same as the control biomolecule, it is determined that the prediction biomolecule and the action gain data meet a first condition.

7. The molecular optimization method according to claim 2, characterized in that, The process of identifying the candidate biomolecule from the plurality of molecular prediction results based on the action gain data includes: For each of the molecular prediction results, a prediction score is determined based on the action gain data. Based on the predicted scores, the candidate biomolecules are determined from the plurality of molecular prediction results.

8. The molecular optimization method according to claim 7, characterized in that, For each of the molecular prediction results, determining the prediction score of the molecular prediction result based on the action gain data includes: For each of the molecular prediction results, determine the number of actions required to transform the biomolecule under investigation into the molecular prediction result; Based on the action gain data and the number of actions, the prediction score of the molecular prediction result is determined.

9. The molecular optimization method according to claim 8, characterized in that, The step of determining the prediction score of the molecular prediction result based on the action gain data and the number of actions includes: Based on the action gain data, a first score is determined; The second score is determined based on the number of actions performed. Based on the first score and the second score, the prediction score of the molecular prediction result is determined.

10. The molecular optimization method according to claim 1, characterized in that, The pre-trained prediction model performs reward predictions on multiple candidate biomolecules to obtain reward prediction results for each of the multiple candidate biomolecules, including: Determine the first molecular sequence information of the first biomolecule and the second molecular sequence information of each of the plurality of candidate biomolecules; Extract a first sequence feature from the first molecular sequence information, and extract a second sequence feature from the second molecular sequence information; Based on the first sequence features and the second sequence features, the prediction model is used to predict rewards and obtain the molecular docking scores of each of the multiple candidate biomolecules with the first biomolecule. Based on the molecular docking score, the reward prediction result for each of the multiple candidate biomolecules is determined.

11. The molecular optimization method according to claim 1, characterized in that, The process of identifying the target biomolecule from the plurality of candidate biomolecules based on the reward prediction results includes: Based on the magnitude of the reward prediction results, the multiple candidate biomolecules are arranged in descending order to obtain the molecular sequence; The first predetermined number of candidate biomolecules in the molecular sequence are identified as the target biomolecule.

12. The molecular optimization method according to claim 1, characterized in that, After identifying the target biomolecule among the plurality of candidate biomolecules, the method further includes: Based on preset verification rules, molecular binding verification is performed between the target biomolecule and the first biomolecule to obtain verification results; In response to the discrepancy between the verification result and the reward prediction result of the target biomolecule, the prediction model is updated with parameters based on the verification result, the target biomolecule, and the first biomolecule.

13. The molecular optimization method according to claim 1, characterized in that, The prediction model is trained in the following way: Multiple sample molecule pairs are obtained, wherein each sample molecule pair includes a first sample biomolecule, a second sample biomolecule, and reference binding data of the first sample biomolecule and the second sample biomolecule; First sample sequence features are extracted from the molecular sequence information of the first sample biomolecule, and second sample sequence features are extracted from the molecular sequence information of the second sample biomolecule. The first sample sequence features and the second sample sequence features are fused into a sample sequence fusion feature; Based on the sample sequence fusion features, the binding affinity of the sample molecule pairs is predicted using the original model to obtain the predicted binding data of the first sample biomolecule and the second sample biomolecule in the sample molecule pairs. Based on the reference combined data and the prediction combined data, the original model is trained to obtain the prediction model.

14. The molecular optimization method according to claim 13, characterized in that, The step of training the original model based on the reference combined data and the prediction combined data to obtain the prediction model includes: For each of the sample pairs, a sub-loss function is determined based on the reference binding data and the predicted binding data; Based on the sub-loss functions of the multiple sample pairs, the total loss function is determined; The original model is trained based on the total loss function to obtain the prediction model.

15. The molecular optimization method according to claim 1, characterized in that, The prediction model includes a first prediction sub-model, a second prediction sub-model, and a third prediction sub-model; The pre-trained prediction model performs reward predictions on multiple candidate biomolecules to obtain reward prediction results for each of the multiple candidate biomolecules, including: For each of the candidate biomolecules, the binding affinity between the candidate biomolecule and the first biomolecule is predicted based on the first prediction sub-model, and a first reward result for the candidate biomolecule is determined based on the binding affinity. The solubility of the candidate biomolecule is predicted based on the second prediction sub-model, and a second reward result for the candidate biomolecule is determined based on the solubility. The toxicity index of the candidate biomolecule is predicted based on the third prediction sub-model, and the third reward result of the candidate biomolecule is determined based on the toxicity index. Based on the first reward result, the second reward result, and the third reward result, the reward prediction result of the candidate biomolecule is determined.

16. The molecular optimization method according to claim 1, characterized in that, The second biomolecule to be optimized was obtained through the following method: Identify multiple reference biomolecules whose binding affinity to the first biomolecule is greater than a preset threshold; Based on the molecular physicochemical information of each of the plurality of reference biomolecules, the second biomolecule is identified among the plurality of reference biomolecules.

17. A molecular optimization device, characterized in that, The molecular optimization device includes: The first determining unit is used to determine the first biomolecule and the second biomolecule to be optimized; An optimization unit is configured to optimize and predict the second biomolecule based on the first biomolecule using a preset molecular optimization model, and determine candidate biomolecules predicted in each round, wherein the candidate biomolecules in each round are predicted based on the candidate biomolecules in the previous round relative to the round. The prediction unit is used to perform reward prediction on multiple candidate biomolecules based on a pre-trained prediction model, and obtain the reward prediction results for each of the multiple candidate biomolecules. The second determining unit is used to determine the target biomolecule among the plurality of candidate biomolecules based on the reward prediction result, wherein the binding affinity between the target biomolecule and the first biomolecule is greater than the binding affinity between the second biomolecule and the first biomolecule.

18. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the molecular optimization method according to any one of claims 1 to 16.

19. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the molecular optimization method according to any one of claims 1 to 16.

20. A computer program product comprising a computer program that is read and executed by a processor of a computer device, causing the computer device to perform the molecular optimization method according to any one of claims 1 to 16.

Images