A method and system for targeted measurement selection of protein variants

Abstract

The application discloses a protein variant target-oriented measurement selection method and system, and belongs to the technical field of computational biology, protein engineering and machine learning assisted experimental design, and comprises the following steps: obtaining a protein variant object, existing measurement data and experimental constraint information, and determining a measurable object set and a target object set according to the information; constructing an association prediction model based on the existing measurement data and a pre-trained model, obtaining an evaluation result of the target object set, prediction uncertainty and a prediction association between candidate measurable objects and the target object, and then calculating a target-oriented measurement value of each candidate measurable object; performing batch redundancy selection under the constraint of an experimental budget based on the value, and obtaining a batch of to-be-measured objects; receiving new measurement results after experimental measurement and updating the association prediction model, and finally identifying a high-performance protein variant from the target object set. The application can effectively improve the measurement design effect of the protein variant under the condition of limited experimental budget.

Description

Technical Field

[0001] This invention belongs to the fields of computational biology, protein engineering, and machine learning-assisted experimental design technology, and particularly relates to a protein variant target-oriented measurement selection method and system. Background Technology

[0002] Protein engineering research typically requires identifying candidate proteins with good function, activity, stability, expression levels, or binding abilities from a large number of protein variants. These protein variants are usually defined relative to the wild-type sequence of the same protein; that is, sequences formed by one or more amino acid substitutions, insertions, deletions, or combinations on the wild-type sequence. Because the number of variants increases rapidly with the number of mutation sites and the degree of mutation, it is impossible to measure all candidates individually. Therefore, it is usually necessary to first measure a small subset of protein variants and then build a predictive model based on the existing measurement data for subsequent identification of high-performance variants.

[0003] In existing technologies, a common approach is to first train a protein variant prediction model using existing measurement data. Then, based on the predicted values, prediction uncertainties, or exploratory value of the candidate variants, the next batch of subjects to be included in the experiment is selected. After this batch of subjects has been measured, new experimental labels are added to the training data, the prediction model is retrained or updated, and the next batch of subjects is selected accordingly. This approach typically forms a multi-round iterative closed loop of "measurement-update-reselection-remeasurement." The selection results of the previous round directly affect the training data and candidate priority of the next round. The entire process often requires multiple trials to gradually approach high-performance variants. The measurement selection focus of this approach is on "which candidate object is more worthwhile to measure." Therefore, its evaluation object is mainly the candidate object itself, rather than whether measuring the object can help identify high-performance variants more quickly in the target object set.

[0004] While the above-mentioned technical approach may work in some experimental scenarios, the following structural problems still exist in actual protein engineering tasks.

[0005] First, existing methods mostly rely on the prediction results of candidate variants themselves for selection, without directly evaluating the improvement effect of measuring the candidate variant on the identification results of high-performance variants in the target object set. Therefore, the objects obtained in this way may not be the most beneficial for subsequent high-performance variant identification.

[0006] Second, existing methods typically assume that the objects currently allowed for measurement and the objects ultimately needed to identify high-performance variants are essentially the same, making it difficult to adapt to experimental design scenarios where the set of measurable objects and the set of target objects are inconsistent. In practical tasks, the set of measurable objects may be limited by library construction methods, mutation order, site selection, experimental costs, and screening throughput, while the set of target objects may be higher-order combinatorial variants, structurally constrained candidates, functionally constrained candidates, or candidate objects generated by computational design processes; the two may not necessarily overlap.

[0007] Third, descriptive information about the target object can be obtained in advance in many tasks, such as the target object sequence, mutation site, structural information, embedding representation, or design source. However, experimental labels for the target object are not yet available during the measurement selection phase. Existing methods typically lack mechanisms to use this type of target object descriptive information to guide measurement selection. Therefore, it is impossible to conduct targeted measurement design around the target object set when labels are unavailable, and it is also difficult to directly lock in high-performance variants after completing a round of target object set evaluation.

[0008] Fourth, while existing multi-round iterative experiments can gradually correct the model, each round requires candidate selection, library construction, experimental measurement, and model updates, which is time-consuming and costly to coordinate. Meanwhile, in single-round batch experiments, if the information duplication among the selected objects is not considered, it is easy to select multiple objects that provide similar evaluation information in the same batch, resulting in intra-batch redundancy, making it difficult to achieve the goal of "evaluating the target object set with a sample in one round and then directly selecting high-performance variants".

[0009] To address the aforementioned problems in the existing technology, there is an urgent need to propose a target-oriented measurement selection method and system for protein variants. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention provides a protein variant target-guided measurement selection method, comprising the following steps: Obtain protein variant objects, existing measurement data, and experimental constraint information; The set of measurable objects is determined based on the experimental constraint information, and the set of target objects is determined based on the research objectives and design requirements; A correlation prediction model is constructed based on at least one of existing measurement data and pre-trained models. The evaluation results and prediction uncertainties of the target object set are obtained through the correlation prediction model, and the predicted correlation between the candidate measurable objects and the objects in the target object set is obtained. Based on the evaluation results, the prediction uncertainty, and the prediction correlation, the target-oriented measurement value of each candidate measurable object is obtained. Based on the target-oriented measurement value, a batch redundancy removal selection is performed under experimental budget constraints to obtain a batch of test objects; Output the batch of objects to be tested, and receive new measurement results after experimental measurement to update the correlation prediction model; Based on the updated association prediction model, high-performance protein variants are identified from the target object set.

[0011] Optionally, in the step of obtaining a protein variant object: the protein variant object is defined relative to the wild-type sequence of the same protein and is represented by at least one of the following descriptive information: mutation site, mutation type, and combination method.

[0012] Optionally, in the step of determining the set of measurable objects based on the experimental constraint information: the set of measurable objects is determined based on at least one of the experimental boundary conditions, construction rules, resource constraints, and measurement constraints.

[0013] Optionally, in the step of determining the target object set according to the research objectives and design requirements: the target object set is determined based on at least one of the following information: research objectives, design constraints, structural constraints, functional constraints, and candidate generation rules, wherein the experimental labels of the objects in the target object set are not used for the selection of the batch of the test objects.

[0014] Optionally, in the step of constructing an association prediction model based on at least one of the existing measurement data and the pre-trained model: The association prediction model can be any one of the following: probabilistic model, kernel method model, Bayesian model, model based on embedding representation, and ensemble model. The association prediction model is trained, initialized, or updated based on at least one of the pre-trained model and existing measurement data; Output the evaluation results of the target object set, the prediction uncertainty of the target object set, and the predicted correlation between the candidate measurable objects and the objects in the target object set.

[0015] Optionally, in the step of obtaining the target-oriented measurement value of each candidate measurable object based on the evaluation results, the prediction uncertainty, and the prediction correlation: The target-oriented measurement value is determined based on at least one of the following indicators: the expected decrease in evaluation uncertainty of the target object set, the expected decrease in high-performance variant identification error, the expected improvement in identification stability, and the expected improvement in high-performance protein variant identification capability.

[0016] Optionally, in the step of performing batch redundancy removal selection under experimental budget constraints based on the target-oriented measurement value to obtain a batch of test objects: Initialize the batch of objects to be tested; Select candidate measurable objects whose target-oriented measurement value meets the preset conditions and add them to the batch of objects to be measured; Based on the impact of the candidate measurable objects already added to the batch of objects to be tested on the evaluation information of the target object set or the high-performance variant identification results, update the target-oriented measurement value of the remaining candidate measurable objects; Repeat the selection and update process until the number of measurements allowed by the experimental budget constraint is reached, thus obtaining a batch of test objects.

[0017] Optionally, the relationship between the set of measurable objects and the set of target objects can be that they are the same, partially overlap, do not overlap, or one contains the other.

[0018] Optionally, when the size of the target object set exceeds a preset threshold, at least one of the following methods can be used to reduce computational costs: random sampling, hierarchical sampling, sampling guided by a protein language model, low-rank representation, block computation, and sparse representation.

[0019] The present invention also provides a protein variant target-guided measurement selection system for implementing the method, comprising: The data acquisition module is used to acquire protein variant objects, existing measurement data, and experimental constraint information; The set determination module is used to determine the set of measurable objects based on the experimental constraint information, and to determine the set of target objects based on the research objectives and design requirements; The association prediction module is used to construct an association prediction model based on at least one of existing measurement data and a pre-trained model, obtain the evaluation results and prediction uncertainty of the target object set through the association prediction model, and obtain the predicted association between the candidate measurable object and the object in the target object set. The measurement value calculation module is used to obtain the target-oriented measurement value of each candidate measurable object based on the evaluation results, the prediction uncertainty, and the prediction correlation. The batch selection module is used to perform batch redundancy removal selection based on the target-oriented measurement value and under experimental budget constraints to obtain a batch of test objects. The output update module is used to output the batch of objects to be tested and to receive new measurement results after the experimental measurement, and update the correlation prediction model. A high-performance variant prediction module is used to identify high-performance protein variants from the target object set based on the updated association prediction model.

[0020] Compared with the prior art, the present invention has the following advantages and technical effects: This invention explicitly distinguishes between the set of measurable objects and the set of target objects during the measurement selection phase. Based on existing measurement data and pre-trained models, it constructs an association prediction model. This model outputs the evaluation results of the target object set, prediction uncertainties, and predicted associations between candidate measurable objects and objects in the target object set. This establishes a computable bridge from measurable objects to target objects, even when experimental labels for target objects are unavailable but descriptive information is available. Furthermore, this invention shifts the core evaluation metric of measurement selection from "whether the candidate object itself is worth measuring" to "whether measuring the object is beneficial for identifying high-performance variants in the target object set." This ensures that the limited experimental budget directly serves the final identification goal, effectively solving the problem of disconnect between measurement selection and identification goals in traditional methods. Simultaneously, under the constraint of a single-round experimental budget, this invention employs a batch redundancy removal selection mechanism. After selecting each candidate measurable object, the marginal value of the remaining candidate objects is updated based on the coverage of the selected object's evaluation information of the target object set. This reduces information duplication among test objects within the same batch, allowing new measurement results to provide more complementary evaluation information and providing a more sufficient data foundation for subsequent effective evaluation of the target object set and direct identification of high-performance variants.

[0021] Furthermore, this invention forms a complete, closed-loop technical chain throughout the entire process, from acquiring protein variant objects and experimental constraint information to ultimately identifying high-performance variants from the target object set. After the experimental measurements of the test objects are completed in batches, this invention can also receive new measurement results and update the associated prediction model, enabling the prediction model to continuously absorb new data and iteratively optimize. It supports both directly identifying high-performance variants after one round of sample evaluation and performing subsequent rounds of test object selection and variant identification when needed, combining flexibility and scalability. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0025] Example 1 This embodiment proposes a protein variant target-oriented measurement selection method. In the measurement selection stage, it explicitly distinguishes between the set of measurable objects and the set of target objects. Utilizing available descriptive information of the target objects and an association prediction model, it calculates the expected improvement of candidate measurable objects on the identification results of high-performance variants in the target object set, and performs batch redundancy removal selection under experimental budget constraints. The technical solution includes the following steps.

[0026] S1. Obtain the protein variant object, existing measurement data, and experimental constraint information.

[0027] The protein variant object includes the protein variant sequence to be analyzed and its corresponding mutation site, mutation type, or other descriptive information. The protein variant object is typically defined relative to the wild-type sequence of the same protein, and the mutation site and mutation type indicate the difference between the object and the wild-type sequence. Existing measurement data includes protein variants for which experimental measurements have been completed and their measurement results; experimental constraints include experimental budget, library design rules, allowed mutation order, measurable range, measurement throughput, or other conditions affecting the feasibility of this round of experiments.

[0028] S2. Determine the set of measurable objects and the set of target objects.

[0029] The measurable set represents the protein variants allowed to be measured under the current experimental conditions; the target set represents the protein variants that ultimately need to be directly identified as high-performance variants. Objects in both the measurable and target sets can be defined relative to the same wild-type sequence, for example, corresponding to single mutations, double mutations, or higher-order combination mutations relative to the wild-type sequence. The measurable and target sets can be identical, partially overlapping, non-overlapping, or one may contain the other. Objects in the target set can provide descriptive information about the target object, but their experimental tags are not used in the selection of test objects in this round.

[0030] For step S2, the set of measurable objects can be determined by predefined experimental boundary conditions, construction rules, resource constraints, measurement constraints, or other information that can limit the range of protein variants allowed to enter the experimental measurement in this round; the set of target objects can be determined by predefined research objectives, design constraints, structural constraints, functional constraints, candidate generation rules, or other information that can limit the range of protein variants of final interest.

[0031] The protein variant object is defined relative to the wild-type sequence of the same protein; the objects in the target object set have available descriptive information during the measurement selection phase, and the experimental labels of the objects in the target object set are not used for the batch deredundancy selection.

[0032] S3. Construct an association prediction model based on a pre-trained model and / or existing measurement data.

[0033] Association prediction models can be built based on pre-trained models, existing measurement data, or a combination of both. They receive descriptive information about the set of measurable objects and the set of target objects, learn the relationship between protein variant descriptions and measurement results, and then output the evaluation results of the target object set, prediction uncertainty, and predicted associations between candidate measurable objects and target objects. For cases where existing measurement data is scarce or even nonexistent, a pre-trained model can be directly used for initial evaluation of the target object set.

[0034] For step S3, the association prediction model can be a probabilistic model, a kernel method model, a Bayesian model, an embedding representation-based model, an ensemble model, or other models that can characterize the predicted association between objects; the protein variant object description information can include sequence, mutation site, mutation order, structural features, protein language model (PLM) embedding, or design source.

[0035] S4. Calculate the target-oriented measurement value of candidate measurable objects.

[0036] For each candidate measurable object in the set of measurable objects, based on the evaluation results, prediction uncertainty and prediction association output by the association prediction model, the expected improvement of the high-performance variant identification results in the target object set after measuring the candidate measurable object is estimated, and the expected improvement is defined as the target-oriented measurement value of the candidate measurable object.

[0037] For step S4, the target-oriented measurement value is determined based on the expected decrease in evaluation uncertainty of the target object set, the expected decrease in high-performance variant identification error, the expected improvement in identification stability, the expected improvement in high-performance protein variant identification capability, or other indicators that can reflect the degree of improvement in high-performance variant identification of the target object set.

[0038] S5. Perform batch redundancy removal selection under experimental budget constraints.

[0039] Based on the target-oriented measurement value output in step S4, one or more candidate measurable objects are selected to form a batch of test objects. After each selection of a candidate measurable object, the marginal value of the remaining candidate measurable objects is updated according to the coverage of the selected object's evaluation information on the target object set, in order to reduce information duplication within the same batch of test objects. The selection and updating process is repeated until the experimental budget constraint is met, thus obtaining the batch of test objects.

[0040] For step S5, batch redundancy removal selection can be achieved through condition updates, marginal value recalculation, diversity constraints, approximate subset selection, or other redundancy removal strategies. The key is that the selected objects will affect the target-oriented measurement value of the remaining candidate objects.

[0041] S6. Output the batch of objects to be tested, and receive new measurement results after the experimental measurement to update the prediction model.

[0042] After the experimental measurements of the batch of test objects are completed, new measurement results are received, and the prediction model is updated based on these new results. These new measurement results can be incorporated into existing measurement data to update the association prediction model. The updated association prediction model can then be used to identify high-performance variants in the target object set or for the next round of test object selection.

[0043] S7. Identify high-performance protein variants in the target object set.

[0044] Based on the evaluation results of the target object set obtained in steps S3 to S6, the prediction uncertainty, the prediction association, and the updated association prediction model, candidate protein variants in the target object set are identified to obtain high-performance protein variants. These high-performance protein variants may be objects with high function, activity, stability, expression level, or binding ability, or other objects that meet the preset research objectives.

[0045] When the target object set is large, computational costs can be reduced through random sampling, hierarchical sampling, protein language model-guided sampling, low-rank representation, block computation, or sparse representation. Protein language model-guided sampling can construct an approximate target object set using target object representation, feasibility score, diversity measure, or sampling weights, but it does not replace the computational logic for goal-oriented measurement value.

[0046] Compared with the prior art, the present invention has at least the following technical effects: This invention explicitly distinguishes between the measurable object set and the target object set during the measurement selection phase, and establishes a predictive correlation between them based on an association prediction model. This ensures that regardless of whether the measurable object set and the target object set are identical, partially overlapping, non-overlapping, or one contains the other, the selection of test objects can be based on the evaluation results and prediction uncertainties of the target object set. This ensures that the limited experimental budget is always allocated around the target object set for which high-performance variants need to be identified, avoiding the waste of measurement resources caused by inconsistent sets. It also makes budget allocation more directly serve the overall process of prior evaluation of the target object set and subsequent identification of high-performance variants.

[0047] This invention addresses the situation where experimental labels for objects in a target object set are unavailable, but descriptive information is available. It outputs the evaluation results of the target object set, prediction uncertainty, and predicted correlations between candidate measurable objects and objects in the target object set through an association prediction model. Based on the above information, the target-oriented measurement value of the candidate measurable objects is obtained. This allows measurement selection to no longer depend on the target object labels, but to complete the measurement design by utilizing the correlation between the target object description information and existing measurement data, thereby achieving target-oriented measurement selection under the condition that labels are unavailable.

[0048] This invention changes the evaluation criteria for measurement selection from the predicted value or uncertainty of the candidate measurable object itself to the expected improvement of the measurement results on the identification results of high-performance variants in the target object set. The test object determined in this way does not only have outstanding performance, but can maximize the identification of high-performance variants in the target object set. This establishes a direct correlation between measurement selection and the final identification target, making up for the technical defect of the separation between the measurement and identification links in traditional methods.

[0049] This invention employs a batch redundancy removal selection mechanism. After selecting each candidate measurable object, the target-oriented measurement value of the remaining candidate measurable objects is updated based on the coverage of the selected objects' evaluation information of the target object set. This allows each object within the same batch of measurable objects to supplement the evaluation information of the target object set from different perspectives, effectively reducing information duplication within the batch. As a result, the maximum complementary evaluation information can be obtained within the budget of a single round of experiment, improving the utilization efficiency of measurement resources.

[0050] This invention can effectively evaluate the target set and directly identify subsequent high-performance variants in a single or few-round experimental process. By batch de-redundancy selection and updating the association prediction model with new measurement results, it reduces the repeated execution of candidate screening, library construction, experimental measurement and model updating caused by multi-round iterative experiments, thereby shortening the experimental cycle and reducing coordination costs. It is especially suitable for protein engineering experimental design scenarios with one-time or few rounds.

[0051] This invention can adapt to protein variants with different mutation orders, the set of measurable objects under different library construction rules, the set of target objects under different design goals, and measurement tasks under different experimental protocol constraints. Through the configurability of the set determination step in step S2 and the compatibility of the association prediction model in step S3 with various variant description information (including sequence, mutation site, mutation order, structural features, protein language model embedding and design source), this invention has good applicability and versatility.

[0052] This invention can be used in conjunction with various association prediction models such as probabilistic models, kernel method models, Bayesian models, embedding representation-based models, and ensemble models. It is also compatible with various protein variant characterization methods. Through the flexible selection of the model construction step in step S3 and the compatibility of the model output type in the goal-oriented measurement value calculation step in step S4, this invention can be seamlessly integrated into the existing computing frameworks of different R&D teams and has good model interface compatibility.

[0053] When the target set is large, this invention can reduce computational costs by using at least one of the following methods: random sampling, hierarchical sampling, sampling guided by protein language models, low-rank representation, block computation, and sparse representation. By introducing the sampling or approximate computation strategy in step S5, the calculation of target-oriented measurement value can maintain the effectiveness of the evaluation while having good scalability, making it suitable for large-scale protein variant screening tasks.

[0054] Example 2 In this embodiment, as Figure 1 As shown, a protein variant target-guided measurement selection method includes the following steps: S1. Obtain the protein variant object, existing measurement data, and experimental constraint information.

[0055] Obtain the protein variant object to be analyzed, which is defined relative to the wild-type sequence of the same protein and represented by mutation site, mutation type, combination method or other descriptive information; obtain existing measurement data, which includes protein variant objects for which experimental measurements have been completed and their corresponding experimental readings; obtain experimental constraint information, which includes experimental budget, number of objects allowed to be measured, library construction range, allowed mutation order, screening throughput, resource limitations or other conditions affecting the feasibility of this round of experiments.

[0056] In this embodiment, researchers first identify a protein system to be optimized and uniformly represent all candidate protein variants as mutations relative to the wild-type sequence of that protein. For example, single-point mutations, multi-point combination mutations, insertion variants, deletion variants, or other designable variant forms can be used to represent different objects. Existing measurement data can come from previous experimental results or from a small subset of measurements already obtained in the same round of research. If existing measurement data is limited or even nonexistent, subsequent steps can still be performed, in which case the pre-trained model provides initial evaluation capabilities. Experimental constraints are pre-defined by the current experimental conditions, such as the maximum number of objects allowed to be measured in this round, the allowed mutation forms, the allowed site regions to be covered, and the acceptable experimental cost and duration.

[0057] S2. Determine the set of measurable objects and the set of target objects.

[0058] The set of measurable objects is determined based on experimental constraints, and the set of target objects is determined based on research objectives and design requirements. The set of measurable objects represents the range of protein variants that are allowed to be measured under the current experimental conditions, while the set of target objects represents the range of protein variants from which we ultimately hope to directly identify high-performance variants.

[0059] In this embodiment, the set of measurable objects is determined by predefined experimental boundary conditions, construction rules, resource constraints, measurement constraints, or other information that limits the range of protein variants allowed to be measured in this round of experiments. For example, this round of experiments may only allow the construction of low-order mutant objects, or only allow combinatorial designs at several key sites; objects meeting these conditions constitute the set of measurable objects. The set of target objects is determined by predefined research objectives, design constraints, structural constraints, functional constraints, candidate generation rules, or other information that limits the range of protein variants of ultimate interest. For example, the set of target objects may be a set of higher-order combinatorial variants, or a set of candidate objects that meet functional objectives, structural stability requirements, or design rules. The set of measurable objects and the set of target objects may be the same, partially overlap, or be different from each other. Objects in the set of target objects can provide target object description information.

[0060] S3. Construct an association prediction model based on a pre-trained model and / or existing measurement data.

[0061] A correlation prediction model is constructed based on a pre-trained model, existing measurement data, or a combination of both. The correlation prediction model receives descriptive information of the set of measurable objects and the set of target objects, learns the relationship between the protein variant descriptive information and the measurement results, and outputs the evaluation results of the set of target objects, the prediction uncertainty, and the predicted correlation between candidate measurable objects and target objects.

[0062] In this embodiment, when existing measurement data is available, the association prediction model is trained or updated using this data, enabling the model to learn the relationship between protein variant description information and experimental readings from the measured objects. When existing measurement data is scarce or even zero, the pre-trained model can be directly used to perform an initial evaluation of the target object set. The pre-trained model can be derived from historical protein variant data, a general protein representation model, or other models capable of providing initial structural characterization, sequence characterization, or variant effect estimation. After receiving the description information of the measurable object set and the target object set, the association prediction model provides an overall evaluation result and corresponding uncertainty estimate for the target object set, and further characterizes the predicted association between candidate measurable objects and target objects. The predicted association is used to indicate the extent to which a new measurement result of a candidate measurable object might change the identification judgment of high-performance variants in the target object set. Through this step, a computable evaluation basis for the target object set can be formed even when the experimental labels for the target objects are not yet available.

[0063] S4. Calculate the target-oriented measurement value of candidate measurable objects.

[0064] For a candidate measurable object in the measurable object set, based on the evaluation results, prediction uncertainty and prediction association output by the association prediction model, the expected improvement of the high-performance variant identification results in the target object set after measuring the candidate measurable object is estimated, and the expected improvement is defined as the target-oriented measurement value of the candidate measurable object.

[0065] In this embodiment, the decision to measure a candidate measurable object is not based solely on its predicted value or uncertainty, but rather on evaluating whether measuring the object will better identify high-performance variants within the target object set. Specifically, if the measurement result of a candidate measurable object is expected to significantly reduce the overall evaluation uncertainty of the target object set or improve the stability of high-performance variant identification, then the candidate measurable object has high target-oriented measurement value. Conversely, if a candidate measurable object performs well, but its measurement result provides limited assistance in identifying high-performance variants within the target object set, then its target-oriented measurement value is low. Through this step, the evaluation object for measurement selection shifts from "whether the candidate object itself is outstanding" to "whether measuring the object is beneficial for faster and more stable identification of high-performance variants within the target object set."

[0066] S5. Perform batch redundancy removal selection under experimental budget constraints.

[0067] Based on the goal-oriented measurement value, candidate measurable objects are selected in batches, and the marginal value of the remaining candidate measurable objects is updated after each selection to reduce information duplication within the batch of test objects, until the experimental budget constraint is met, and the batch of test objects is obtained.

[0068] In this embodiment, the candidate measurable object with the highest current target-oriented measurement value is first selected from the set of measurable objects and added to the batch of objects to be measured. Then, based on the coverage of the selected object's evaluation information on the target object set, the marginal contribution of the remaining candidate measurable objects is re-evaluated. If some remaining candidate objects provide highly similar evaluation information to the selected objects on the target object set, their marginal value decreases; if some remaining candidate objects can supplement information from different regions, different variant types, or different sources of uncertainty, their marginal value remains high. The process of "selecting one object—updating the marginal value of remaining objects" is repeated until the number of measurements allowed by the experimental budget is reached. Through this batch redundancy-removal selection process, the objects in the batch of objects to be measured can provide complementary information as much as possible, thereby completing an effective evaluation of the target object set with fewer measurement objects and providing a more favorable data foundation for directly identifying high-performance variants.

[0069] S6. Output the batch of objects to be tested, and receive new measurement results after the experimental measurement to update the prediction model.

[0070] Output the batch of test objects to the subsequent experimental process; after the batch of test objects completes the experimental measurement, receive the new measurement results and merge them into the existing measurement data; when necessary, update the correlation prediction model based on the updated existing measurement data.

[0071] S7. Identify high-performance protein variants in the target object set.

[0072] Based on the evaluation results of the target object set, prediction uncertainty, prediction association, and updated association prediction model obtained in steps S3 to S6, candidate protein variants in the target object set are identified to obtain high-performance protein variants.

[0073] In this embodiment, the batch of test objects can be exported in an experimentally executable format, such as a variant sequence list, mutation numbering table, library construction list, or an input file recognizable by an automated experimental platform. After the experimental measurements are completed, new measurement results are incorporated into existing measurement data and can be used to update the association prediction model. Through the aforementioned steps, this batch of test objects in this embodiment is first used to complete an effective evaluation of the target object set, allowing researchers to directly identify high-performance variants in the target object set; when further iterations are needed, the updated association prediction model can also be used for the selection of test objects in subsequent batches.

[0074] Through the above steps, this embodiment achieves a technical process of selecting a small number of test objects from the set of testable objects, evaluating the target object set first, and then directly identifying high-performance variants, even when the experimental labels for the target objects are unavailable but the descriptive information of the target objects is available. This process not only reduces reliance on multiple rounds of experimental feedback but also improves the utilization efficiency of limited experimental budgets in protein variant engineering tasks. New measurement results returned after completing experimental measurements can also be used to update the association prediction model to support target-oriented measurement selection in subsequent batches. Step S7 is used to directly identify high-performance protein variants in the target object set based on the evaluation of the target object set.

[0075] like Figure 2 As shown in the figure, this embodiment also provides a protein variant target-oriented measurement selection system, including: a data acquisition module, a set determination module, an association prediction module, a measurement value calculation module, a batch selection module, an output update module, and a high-performance variant prediction module.

[0076] The data acquisition module acquires protein variant objects, existing measurement data, and experimental constraint information. The set determination module determines the set of measurable objects based on experimental constraints and the set of target objects based on the research task. The association prediction module constructs an association prediction model based on a pre-trained model and / or existing measurement data, and outputs the target object set evaluation results, prediction uncertainty, and predicted associations between candidate measurable objects and target objects. The measurement value calculation module calculates the target-oriented measurement value of candidate measurable objects based on the output of the association prediction model. The batch selection module performs batch redundancy removal selection under experimental budget constraints and outputs the batch of test objects. The output update module outputs the batch of test objects and receives new measurement results after experimental measurement to update the association prediction model. The high-performance variant prediction module identifies high-performance protein variants in the target object set based on the target object set evaluation results, prediction uncertainty, and the updated association prediction model.

[0077] Furthermore, after each selection of candidate measurable objects, the batch selection module updates the target-oriented measurement value of the remaining candidate measurable objects based on the impact of the selected objects on the evaluation information of the target object set or the high-performance variant identification results; the output update module receives the new measurement results and triggers the association prediction module to update, so as to form an iteratively updated target-oriented measurement selection closed-loop process.

[0078] It should be understood that the protein variant target-guided measurement selection system provided in this embodiment of the invention possesses all the advantages of the protein variant target-guided measurement selection method provided in the above embodiments.

[0079] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A protein variant target-guided measurement selection method, characterized in that, Includes the following steps: Obtain protein variant objects, existing measurement data, and experimental constraint information; The set of measurable objects is determined based on the experimental constraint information, and the set of target objects is determined based on the research objectives and design requirements; A correlation prediction model is constructed based on at least one of existing measurement data and pre-trained models. The evaluation results and prediction uncertainties of the target object set are obtained through the correlation prediction model, and the predicted correlation between the candidate measurable objects and the objects in the target object set is obtained. Based on the evaluation results, the prediction uncertainty, and the prediction correlation, the target-oriented measurement value of each candidate measurable object is obtained. Based on the target-oriented measurement value, a batch redundancy removal selection is performed under experimental budget constraints to obtain a batch of test objects; Output the batch of objects to be tested, and receive new measurement results after experimental measurement to update the correlation prediction model; Based on the updated association prediction model, high-performance protein variants are identified from the target object set.

2. The method according to claim 1, characterized in that, In the step of obtaining a protein variant object: the protein variant object is defined relative to the wild-type sequence of the same protein and is represented by at least one of the following descriptive information: mutation site, mutation type, and combination method.

3. The method according to claim 1, characterized in that, In the step of determining the set of measurable objects based on the experimental constraint information: the set of measurable objects is determined based on at least one of the experimental boundary conditions, construction rules, resource constraints, and measurement constraints.

4. The method according to claim 1, characterized in that, In the step of determining the target object set according to the research objectives and design requirements: the target object set is determined based on at least one of the following information: research objectives, design constraints, structural constraints, functional constraints, and candidate generation rules. The experimental labels of the objects in the target object set are not used for the selection of the batch of the test objects.

5. The method according to claim 1, characterized in that, In the step of constructing an association prediction model based on at least one of the existing measurement data and pre-trained models: The association prediction model can be any one of the following: probabilistic model, kernel method model, Bayesian model, model based on embedding representation, and ensemble model. The association prediction model is trained, initialized, or updated based on at least one of the pre-trained model and existing measurement data; Output the evaluation results of the target object set, the prediction uncertainty of the target object set, and the predicted correlation between the candidate measurable objects and the objects in the target object set.

6. The method according to claim 1, characterized in that, In the step of obtaining the target-oriented measurement value of each candidate measurable object based on the evaluation results, the prediction uncertainty, and the prediction correlation: The target-oriented measurement value is determined based on at least one of the following indicators: the expected decrease in evaluation uncertainty of the target object set, the expected decrease in high-performance variant identification error, the expected improvement in identification stability, and the expected improvement in high-performance protein variant identification capability.

7. The method according to claim 1, characterized in that, Based on the aforementioned target-oriented measurement value, and under experimental budget constraints, the step of performing batch redundancy removal selection to obtain a batch of test objects includes: Initialize the batch of objects to be tested; Select candidate measurable objects whose target-oriented measurement value meets the preset conditions and add them to the batch of objects to be measured; Based on the impact of the candidate measurable objects already added to the batch of objects to be tested on the evaluation information of the target object set or the high-performance variant identification results, update the target-oriented measurement value of the remaining candidate measurable objects; Repeat the selection and update process until the number of measurements allowed by the experimental budget constraint is reached, thus obtaining a batch of test objects.

8. The method according to claim 1, characterized in that, The relationship between the set of measurable objects and the set of target objects is that they are the same, partially overlap, do not overlap, or one contains the other.

9. The method according to claim 1, characterized in that, When the size of the target object set exceeds a preset threshold, at least one of the following methods is used to reduce computational cost: random sampling, hierarchical sampling, sampling guided by a protein language model, low-rank representation, block computation, and sparse representation.

10. A protein variant target-guided measurement selection system, characterized in that, For implementing the method according to any one of claims 1-9, comprising: The data acquisition module is used to acquire protein variant objects, existing measurement data, and experimental constraint information; The set determination module is used to determine the set of measurable objects based on the experimental constraint information, and to determine the set of target objects based on the research objectives and design requirements; The association prediction module is used to construct an association prediction model based on at least one of existing measurement data and a pre-trained model, obtain the evaluation results and prediction uncertainty of the target object set through the association prediction model, and obtain the predicted association between the candidate measurable object and the object in the target object set. The measurement value calculation module is used to obtain the target-oriented measurement value of each candidate measurable object based on the evaluation results, the prediction uncertainty, and the prediction correlation. The batch selection module is used to perform batch redundancy removal selection based on the target-oriented measurement value and under experimental budget constraints to obtain a batch of test objects. The output update module is used to output the batch of objects to be tested and to receive new measurement results after the experimental measurement, and update the correlation prediction model. A high-performance variant prediction module is used to identify high-performance protein variants from the target object set based on the updated association prediction model.

Images