A protein sequence closed-loop evolution design method and system based on a self-evolution multi-agent system and application thereof

Abstract

The application relates to the field of artificial intelligence assisted protein engineering, and discloses a protein sequence closed-loop evolution design method and system based on a self-evolution multi-agent system. The system comprises an arranger, an evaluator, a special-purpose sub-agent, a product registration module, a memory system and a downstream biological laboratory, and is used for generating, reviewing and replanning a protein design workflow. A reinforcement learning sub-agent adopts a GDPO-RepKL strategy, each reward component of a candidate sequence in the same group is normalized and weighted to form a total advantage, after batch secondary normalization, a sequence generator is updated based on a pruning proxy loss, a Kullback-Leibler divergence of a constraint strategy model and a reference model distance and a repulsive Jensen-Shannon divergence between candidate sequences, so that mode collapse is reduced and sequence space exploration is maintained. The system is used for data-driven protein redesign, de novo design of binding proteins, de novo design of antibodies and experiment data guided optimization and other protein design tasks, and structured experiment feedback is used for the next round of design.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted protein engineering technology, specifically to a protein sequence closed-loop evolutionary design method, system, and application based on a self-evolving multi-agent system. The self-evolving multi-agent system described in this invention can be used for protein sequence generation, protein function optimization, binding protein design, antibody design, and wet experimental data-guided protein design iterations. The system forms a reflection-evolutionary loop through an arranger, evaluator, multiple specialized sub-agents, and a memory system, enabling the sequence generator to perform task-specific adaptation and continuous optimization in different tasks. This invention is applicable to scenarios such as data-driven protein redesign, de novo binding protein design, de novo antibody design, and wet experimental data-guided antibody design, and is particularly suitable for generating and screening candidate sequences in protein engineering tasks with insufficient structural data, a large candidate space, or requiring multi-objective optimization. Background Technology

[0002] Protein design has significant applications in therapeutic proteins, antibody drugs, enzyme engineering, and synthetic biology. Traditional protein engineering typically relies on experimental procedures such as random mutation, directed evolution, phage display, yeast display, or cell screening. While these methods can yield functional molecules, they often require the construction of large-scale libraries and multiple rounds of screening, resulting in long cycles, high costs, significant trial-and-error, and limited coverable sequence space. In recent years, artificial intelligence methods have been widely used in protein design. Structural information design methods can utilize three-dimensional structures, framework constraints, or interface geometry information for design, exhibiting high biophysical accuracy in certain tasks. However, high-quality structural data is significantly scarce compared to sequence data, and structure prediction, structure sampling, and structure optimization are usually computationally expensive, making it difficult to support large-scale, high-throughput, and rapid iterative candidate sequence generation (Anishchenko et al., Nature 2021, 600, 547-552). Compared to structural information design methods, sequence generation methods can utilize large-scale protein sequence databases, offering advantages such as large training data volume, fast inference speed, and strong scalability. However, simple sequence generation methods often lack explicit structural and functional constraints, leading to insufficient accuracy in tasks such as target-specific binding, antibody heavy-light chain pairing, binding interface optimization, and exploitability control (Ferruz and Höcker, Nat. Mach. Intell. 2022, 4, 521-532.). Reinforcement learning methods can optimize sequence generators towards specific target properties, making them suitable for multi-objective tasks such as functional score, binding interface quality, structural confidence, exploitability, and diversity. However, existing reinforcement learning protein design methods often encounter a trade-off between reward maximization and sequence diversity preservation. Some methods can improve reward scores but cause generated sequences to concentrate in a few similar regions, resulting in pattern collapse; other methods, while maintaining a certain level of diversity, struggle to fully optimize the reward objective (Angermueller et al., ICLR 2020.).

[0003] Furthermore, existing automated protein design workflows are mostly static pipelines, typically involving pre-defined steps for data processing, model training, reward function construction, and candidate evaluation. These workflows lack task-oriented automated construction capabilities, mechanisms for reviewing intermediate results, mechanisms for failure diagnosis and replanning, and mechanisms for uniformly registering training products, reward models, reward functions, and historical experience for reuse in subsequent steps. Therefore, when the task expands from data-driven protein redesign to de novo binding protein design or de novo antibody design, the workflow often needs to be rebuilt, lacking versatility and sustainable evolution capabilities. For antibody and binding protein design, candidate sequences also need to consider developmental indicators such as binding ability, structural stability, heavy-light chain pairing quality, solubility, aggregation risk, and therapeutic antibody profile. Relying solely on a single prediction model or reward function can easily lead to candidate sequences improving on one indicator while degrading on others. Existing methods also lack mechanisms for stably converting wet experimental data, deep mutation scanning data, or screening experimental data into feedback for the next round of computational design.

[0004] Therefore, there is an urgent need for a self-evolving multi-agent system for protein engineering that can maintain the scalability of sequence generation methods while achieving unified, reliable, and sustainable optimization of closed-loop protein sequence design through reflection-evolution loop, multi-agent cooperation, evaluator review, reward model quality control, reward function assembly, reinforcement learning evolution, structure and exploitability evaluation, and wet experimental data-guided feedback. Summary of the Invention

[0005] The purpose of this invention is to provide a protein sequence computation and experimental closed-loop evolutionary design system, method, and application based on a self-evolving multi-agent system. This addresses problems in existing protein design methods, such as high costs of structural data acquisition and computation, insufficient joint constraints on function and structure during sequence generation, difficulty in adapting single models to various protein design tasks, pattern collapse in reinforcement learning optimization, lack of quality control in reward models and functions, reliance on manual coordination between different computational tools, and the disconnect between computational design and biological experimental verification. This invention organizes heterogeneous computing resources, self-evolving multi-agent computational design, biological experimental verification, and experimental data feedback into a unified technical system. This enables protein design tasks to sequentially complete computational planning, candidate sequence generation, multi-objective optimization, quality review, candidate protein preparation, property detection, experimental feedback structuring, and the next round of redesign, thus forming a closed-loop evolutionary process where computational design acts on physical experiments, and experimental measurement results then in turn act on computational design.

[0006] To achieve the above objectives, the present invention provides a protein sequence computation and experimental closed-loop evolutionary design system based on a self-evolving multi-agent system. The system includes a cloud computing and supercomputing platform, a self-evolving multi-agent computational design subsystem, a biological verification laboratory subsystem, and an experimental feedback interface. The cloud computing and supercomputing platform provides the computing and data resources required for model training, sequence generation, reinforcement learning, structure prediction, molecular docking, candidate evaluation, and data storage. The self-evolving multi-agent computational design subsystem receives protein design tasks, generates and reviews workflows, schedules multiple specialized sub-agents to execute the workflows, and outputs candidate protein sequences and corresponding experimental task information. The biological validation laboratory subsystem receives candidate protein sequences and experimental task information, prepares corresponding candidate proteins, binding proteins, or antibodies, and obtains experimental measurement results related to the function, structure, binding ability, stability, or exploitability of the candidate proteins. The experimental feedback interface transmits experimental task information and experimental measurement results between the self-evolving multi-agent computational design subsystem and the biological validation laboratory subsystem, establishes a correspondence between candidate protein sequences, generation rounds, computational evaluation results, experimental samples, and experimental measurement results based on candidate numbers, and transforms the experimental measurement results into structured feedback data for the next round of computational design.

[0007] The cloud computing and supercomputing platform includes a communication unit, a storage unit, a computing unit, an input / output interface, a task scheduling unit, and an execution isolation environment. The computing unit includes at least one of central processing unit (CPU) computing resources and graphics processing unit (GPU) computing resources, used to perform at least one computational task among adaptive pre-training, supervised fine-tuning, reward model training, reinforcement learning optimization, protein structure prediction, protein complex structure prediction, molecular docking, and candidate sequence evaluation. The storage unit is used to store protein sequence data, experimental data, model parameters, workflows, training configurations, intermediate products, candidate protein sequences, structure prediction results, experimental task information, and structured feedback data. The communication unit and input / output interface are used to transmit tasks, products, candidate sequences, and experimental results between the cloud computing and supercomputing platform, the self-evolving multi-agent computational design subsystem, and the biological verification laboratory subsystem. The task scheduling unit allocates computing resources according to task resource type, task dependencies, and task execution status; the execution isolation environment executes training programs, structure prediction programs, evaluation programs, or external tools in a runtime environment isolated from the main control process, and returns training progress, execution status, log information, and product references to the self-evolving multi-agent computational design subsystem.

[0008] The self-evolving multi-agent computational design subsystem includes a task input module, an orchestrator, an evaluator, a task construction sub-agent, a data processing sub-agent, a task adaptive pre-training sub-agent, a supervised fine-tuning sub-agent, a reward model training and evaluation sub-agent, a reward function assembly sub-agent, a reinforcement learning sub-agent, an evaluation sub-agent, a product registration module, a memory system, and a candidate sequence output interface. The task input module receives a protein design task, and at least one of the following: a protein sequence, a target sequence, an antigen sequence, experimental data, and design constraints corresponding to the protein design task. The protein design task includes at least one of data-driven protein redesign, de novo binding protein design, de novo antibody design, and experimental data-guided protein sequence optimization. The self-evolving multi-agent computational design subsystem selects appropriate dedicated sub-agents and forms an executable workflow based on the task type, available data, registered products, computational resources, and historical execution records.

[0009] The orchestrator is used to parse protein design tasks, generate a workflow containing one or more typified task items based on task objectives, input data, registered products, and historical records in the memory system, and schedule corresponding dedicated sub-agents according to the dependencies between the typified task items. Each typified task item includes at least a task type, task instructions, structured input, expected output, task reason, resource type, and dependencies. The task type includes at least one of task construction, data processing, task adaptive pre-training, supervised fine-tuning, reward model training and evaluation, reward function assembly, reinforcement learning evolution, candidate evaluation, and task termination. The task reason is used to record the objective or basis for setting the corresponding task item, and the resource type is used to represent the central processing unit, graphics processing unit, storage space, external tools, or experimental resources required to execute the corresponding task item. When a task item fails to execute, the upstream product it depends on is unavailable, the product type does not match the downstream task, or the evaluation result does not meet the corresponding conditions, the orchestrator modifies the task item, inserts a repair task, adjusts the task order, or regenerates the subsequent workflow based on the correction information output by the evaluator.

[0010] The evaluator is used to review the workflow, structured input, task dependencies, intermediate products, candidate protein sequences, and task execution results before and after the execution of corresponding typified task items. The review includes rule review and semantic review. Rule review includes checking at least one of the following: whether the task item conforms to a preset structured input protocol; whether the data path or model path is accessible; whether the upstream product is registered; whether the product type matches the downstream task; whether the data column role is clearly defined; whether the task execution order satisfies dependencies; whether the reward model meets quality access criteria; whether the reward function passes procedural testing; and whether repetitive tasks need to be executed. Semantic review includes determining at least one of the following: whether the selected specialized sub-agent matches the protein design goal; whether the data field role matches the task type; whether the optimization direction of the reward component is consistent with the task goal; whether the parameter settings meet the confirmed scientific constraints; and whether the current workflow should continue, be modified, terminated, or replanned. If the review fails, the evaluator outputs structured correction information including the review status, error fields, error reasons, repair suggestions, and original workflow information. The orchestrator modifies or replans the workflow based on this structured correction information.

[0011] The task-building sub-agent is used to transform user-input protein design goals into executable structured task descriptions. These design goals include at least one of the following: improving functional activity, increasing expression levels, improving solubility, improving stability, improving target binding ability, improving binding interface quality, increasing structural confidence, reducing aggregation risk, improving antibody developability, improving candidate sequence novelty, and maintaining candidate sequence diversity. The structured task description includes at least one of the following: task type, input sequence, target properties, candidate length range, chain type, fixed pairing information, evaluation metrics, experimental detection items, and task termination conditions. When there is ambiguity in the antigen sequence, fixed pairing sequence, candidate length range, antibody chain format, target label, evaluation metrics, or experimental detection items, the task-building sub-agent generates structured confirmation information and, after the corresponding scientific parameters are confirmed, enables subsequent training, reinforcement learning, or experimental tasks to enter an executable state.

[0012] The data processing sub-agent performs at least one of the following processing operations on the input protein sequence data, experimental data, or evaluation data: standardization, cleaning, deduplication, amino acid validity checking, column role identification, label data type conversion, label dimension conversion, sequence length statistics, and dataset partitioning, and generates data semantic information. The data semantic information includes at least one of the following: data source, sample quantity, protein sequence field, receptor sequence field, binding protein sequence field, antibody heavy chain sequence field, antibody light chain sequence field, label field, label type, chain type, fixed pairing sequence, candidate length window, sequence length distribution, task scope, and parsing status. The data semantic information is stored as a structured data object and serves as input for task adaptive pre-training, supervised fine-tuning, reward model training, reward function assembly, reinforcement learning optimization, and experimental feedback processing. When the parsing status of the data semantic information indicates that confirmation is needed, the orchestrator only allows the invocation of the task construction sub-agent or the data processing sub-agent to supplement the information, and does not allow direct invocation of the training or reinforcement learning sub-agents that depend on the data semantic information.

[0013] The task-adaptive pre-training sub-agent and the supervised fine-tuning sub-agent are used to adapt the basic protein sequence generator to a task. The basic protein sequence generator is an autoregressive generation model, a protein language model, a sequence generation model, or a combination thereof, capable of generating amino acid sequences based on task input. The task-adaptive pre-training enables the basic protein sequence generator to learn sequence distributions related to the target protein, target protein family, target antigen, or target dataset; the supervised fine-tuning enables the basic protein sequence generator to learn task conditions, target information, antigen information, or the correspondence between the input sequence and the target protein sequence. After task adaptation, a generator product is generated, which includes at least one of the following: model configuration, vocabulary, initialized model parameters, post-trained model parameters, training input template, chain format, sequence length statistics, and maximum sequence length information. The generator product is registered and then available for use by the reinforcement learning sub-agent.

[0014] The reward model training and evaluation sub-agent is used to train single-chain reward models based on protein sequences and their functional tags, or to train multi-chain reward models based on receptor sequences, binding protein sequences, and their tags. The reward model outputs evaluation signals related to functional activity, expression capacity, binding capacity, structural rationality, stability, or exploitability. After training, the reward model training and evaluation sub-agent determines whether it meets the quality admission criteria based on preset evaluation metrics on the validation or test set. For continuous tags, the preset evaluation metrics include at least one of correlation coefficient, rank correlation coefficient, mean squared error, and mean absolute error; for discrete tags, the preset evaluation metrics include at least one of accuracy, precision, recall, F1 score, and area under the curve. Reward models that do not meet the quality admission criteria are not entered into the reward model registry, are not considered qualified reward models for the reward function assembly sub-agent to call, and the evaluator outputs the corresponding failure status and evaluation result.

[0015] The reward function assembly sub-agent is used to assemble multi-objective reward functions based on the reward intent structure, protein design goals, registered reward models, structure evaluation results, experimental data, and design constraints. The multi-objective reward functions include at least one of amino acid validity gating, length gating, position template gating, reward model scoring, fixed pairing rewards, structure proxy rewards, sequence naturalness rewards, diversity rewards, antibody structure funnels, and exploitability proxies. The reward function assembly process includes reward intent review, component parsing, plug-in combination, input / output field checking, procedural testing, and output registration. The reward intent review checks whether the optimization direction of each reward component is consistent with the protein design goals; the input / output field checking checks whether the input and output fields of the reward function are consistent with the preset fields of the corresponding calling interface; the procedural testing checks whether each reward component can generate outputs conforming to preset data types and numerical ranges based on the input candidate sequences. Reward functions that fail the reward intent review, input / output field checking, or procedural testing are not entered into the reward function registry and are not available for the reinforcement learning sub-agent to call, thereby avoiding the use of simulated, placeholder, or failure-backoff reward signals to perform candidate sequence optimization.

[0016] The product registration module includes a generator product registry, a reward model registry, and a reward function registry. Each entry in the generator product registry includes at least one of the following: source sub-agent, output directory, model configuration file, vocabulary file, initialization model parameters, post-trained model parameters, training input template, chain format, sequence length statistics, and maximum sequence length information. Each entry in the reward model registry includes at least one of the following: reward model parameters, task type, chain type, target label, validation or test set evaluation metrics, quality admission status, and source output directory. Each entry in the reward function registry includes at least one of the following: reward function construction program, reward function parameters, reward component weights, output directory, name of the reward component that passed the test, and test report. The product registration module only returns product references whose registration and quality status meet the corresponding calling conditions to downstream typed task items; the reward function assembly sub-agent only calls registered reward models that meet the quality admission conditions; and the reinforcement learning sub-agent only calls registered generator products and reward functions that have passed the test.

[0017] The reinforcement learning sub-agent invokes a registered protein sequence generator and reward function to perform task-specific optimization on the protein sequence generator. The generator generates candidate groups containing multiple candidate protein sequences for the same task input. The reward function calculates one or more reward components for each candidate protein sequence and determines an advantage for policy updates based on these reward components. The adjustment of the advantage or policy update includes at least one of the following methods: determining a group-level diversity reward based on the repetition of candidate protein sequences within the candidate group, and combining the advantage corresponding to the group-level diversity reward with the advantages corresponding to other reward components; scaling the advantage based on the diversity score of the candidate protein sequences and the positive or negative direction of its advantage; dynamically adjusting the pruning threshold of the policy probability ratio based on the repetition of candidate protein sequences; and constructing an exclusionary regularization term based on the divergence between the word-level policy probability distributions of different candidate protein sequences within the candidate group. These adjustment methods can be used individually or in combination to mitigate pattern collapse and maintain the ability to explore the protein sequence space. The reward components include at least one of the following: functional prediction reward, sequence naturalness reward, amino acid legitimacy reward, length constraint reward, structure confidence reward, binding interface reward, fixed pairing reward, diversity reward, solubility-related reward, and exploitability reward.

[0018] Furthermore, the reinforcement learning sub-agent calculates the Kullback-Leibler divergence between the current policy model and the reference model to limit the current policy model from deviating excessively from the protein sequence distribution represented by the reference model due to reward optimization. For candidate sequence pairs within the same candidate group, at the common effective generation positions of the two candidate protein sequences, the Jensen-Shannon divergence between the word conditional probability distributions output by the current policy model under the individual generation conditions of the two candidate protein sequences is calculated. The word-level Jensen-Shannon divergences at the common effective generation positions are aggregated to obtain the sequence pair-level Jensen-Shannon divergence, and the candidate group-level exclusionary Jensen-Shannon divergence is obtained based on the sequence pair-level Jensen-Shannon divergences of multiple candidate sequence pairs within the candidate group. The candidate group-level exclusionary Jensen-Shannon divergence is not used as a normal reward component in intra-group normalization and secondary normalization, but rather participates in policy updates at the policy loss layer in a direction that promotes the increase of probability distribution differences between different candidate generation paths. By adjusting the distribution constraints between the current strategy model and the reference model, as well as the distribution differences between different generation paths within the candidate group, the degree to which candidate sequences are concentrated in a few similar generation patterns is reduced while maintaining the optimization of the target reward.

[0019] The evaluation sub-agent is used to perform at least one of the following evaluations on candidate protein sequences output by the reinforcement learning sub-agent or the protein sequence generator: functional evaluation, sequence evaluation, structural evaluation, binding interface evaluation, and exploitability evaluation. The evaluation includes at least one of the following: functional prediction score, sequence naturalness, hydrophobicity or solubility, global structure confidence, residue-by-residue structure confidence, inter-chain interface confidence, interface prediction alignment error, antibody heavy chain / light chain pairing quality, sequence novelty, sequence repetition ratio, surface hydrophobicity, surface positive charge, surface negative charge, charge symmetry, and complementarity-determining region length. Based on one or more evaluation results, the evaluation sub-agent performs legitimacy filtering, length filtering, repetitive sequence filtering, quality ranking, and experimental validation priority ranking on the candidate protein sequences, and generates an experimental candidate list including candidate number, candidate protein sequence, generation round, reward component, calculated evaluation result, filtering status, and experimental validation priority.

[0020] The memory system is used to store at least one of the following: user tasks, workflows, categorized task items, evaluator review results, execution results of dedicated sub-agents, intermediate product references, candidate protein sequences, reward components, candidate evaluation results, experimental feedback data, error fields, failure reasons, and effective optimization strategies. The memory system includes short-term memory units, memory compression units, insight memory units, and long-term memory units. The short-term memory units store the workflow, execution records, candidate sequences, and evaluation results of the current design round. The memory compression unit generates compressed records from the short-term memory, including task type, execution status, evaluation metrics, error fields, failure reasons, and processing actions, when the short-term memory record size or vocabulary size reaches a preset context capacity. The insight memory unit stores the filtered compressed records. The long-term memory unit stores cross-task historical records and returns historical records matching the current protein design task to the orchestrator and evaluator through a retrieval interface. The orchestrator generates the next round's workflow based on the historical records, the current round's evaluation results, and experimental feedback data, thus forming a reflective-evolutionary loop of generation, evaluation, memory, and replanning.

[0021] The candidate sequence output interface is used to transmit the list of experimental candidates and corresponding experimental task information to the biological validation laboratory subsystem. The experimental task information includes at least one of the following: candidate number, candidate protein sequence, protein type, strand format, target target or antigen, suggested expression system, suggested purification method, experimental indicators to be tested, sample requirements, and experimental validation priority. The candidate protein sequence and experimental task information are output in the form of sequence files, data tables, database records, experimental task sheets, or other data formats readable by the biological validation laboratory subsystem. The candidate number remains associated throughout the computational design, nucleic acid or protein preparation, expression and purification, property detection, and feedback input processes to establish a traceable correspondence between the computationally generated object, physical experimental samples, and experimental measurement results.

[0022] The biological validation laboratory subsystem includes at least a candidate protein preparation unit and a property detection unit. The candidate protein preparation unit is used to obtain corresponding candidate protein experimental samples based on candidate protein sequences, and includes at least one of a sequence synthesis unit, a vector construction unit, a protein expression unit, and a protein purification unit. The sequence synthesis unit is used to obtain the nucleic acid sequence encoding the corresponding candidate protein or directly obtain the corresponding amino acid sequence based on the candidate protein sequence; the vector construction unit is used to construct the nucleic acid sequence encoding the candidate protein into an expression vector; the protein expression unit is used to express the candidate protein in prokaryotic cells, eukaryotic cells, cell-free expression systems, or other protein expression systems; the protein purification unit is used to separate and purify the expression product. The property detection unit is used to detect the candidate protein experimental samples according to the experimental task information and obtain at least one experimental measurement result that can characterize the candidate protein's function, binding ability, stability, or exploitability.

[0023] The property detection unit includes at least one of the following: expression level detection unit, purification yield detection unit, functional activity detection unit, binding detection unit, affinity detection unit, stability detection unit, aggregation detection unit, cell binding detection unit, immunogenicity-related detection unit, and exploitability detection unit. The binding and affinity detection can employ enzyme-linked immunosorbent assay (ELISA), biolayer interferometry (BII), surface plasmon resonance (SPR), cell binding detection, or other detection methods capable of obtaining binding signals or affinity parameters between the candidate protein and the target. The stability and exploitability detection can include thermal stability detection, size exclusion chromatography (SLC), aggregation ratio detection, surface hydrophobicity detection, charge-related detection, or therapeutic antibody profiling. The experimental measurement results include at least one of the following: expression level, purification yield, functional activity, binding signal, affinity parameter, binding rate, dissociation rate, stability index, aggregation ratio, cell binding signal, immunogenicity-related index, and exploitability index.

[0024] For data-driven protein redesign tasks, the biological validation laboratory subsystem is used to detect candidate proteins for expression, purification, functional activity, stability, solubility, or other target properties. For de novo binding protein design tasks, the biological validation laboratory subsystem is used to detect candidate binding proteins for at least one of the following: expression, purification, target binding, binding affinity, functional activity, stability, and immunogenicity risk. For de novo antibody design or data-driven antibody optimization tasks, the biological validation laboratory subsystem is used to detect candidate antibodies for at least one of the following: expression, purification, antigen binding, binding affinity, selectivity, stability, aggregation tendency, and exploitability. The biological validation laboratory subsystem selects appropriate experimental samples and detection items based on the experimental candidate list and returns the candidate number, batch number, and experimental measurement results of the experimental samples to the self-evolving multi-agent computational design subsystem via the experimental feedback interface.

[0025] The experimental feedback interface includes an experimental task output interface, an experimental result receiving interface, and a feedback data conversion unit. The experimental task output interface sends an experimental candidate list and experimental task information to the biological validation laboratory subsystem. The experimental result receiving interface receives experimental measurement results output by the biological validation laboratory subsystem. The feedback data conversion unit associates the experimental measurement results with candidate protein sequences, generation rounds, reward components, calculated evaluation results, experimental samples, and experimental batches based on the candidate number, and converts the experimental measurement results into structured feedback data with preset fields. The structured feedback data includes at least one of the following: candidate number, protein sequence, antibody heavy chain sequence, antibody light chain sequence, antigen or target identifier, generation round, expression level, purification yield, binding signal, affinity parameter, binding rate, dissociation rate, stability index, aggregation ratio, cell binding signal, functional activity, experimental batch, experimental failure reason, and experimental label.

[0026] The structured feedback data is written to the storage unit via the communication unit and input / output interface of the cloud computing and supercomputing platform, and input to at least one of the following: a data processing sub-agent, a reward model training and evaluation sub-agent, a reward function assembly sub-agent, an evaluation sub-agent, or a memory system. The data processing sub-agent performs field identification, data type conversion, label standardization, quality checks, and dataset partitioning on the structured feedback data. The reward model training and evaluation sub-agent trains or updates the reward model based on the processed structured feedback data. The reward function assembly sub-agent forms or adjusts experiment-related reward components based on experimental measurement results. The evaluation sub-agent adjusts candidate filtering conditions or experimental verification priorities based on experimental results. The memory system stores successful experiment records, failed experiment records, and their correspondence with candidate sequences.

[0027] When the quantity and quality of the structured feedback data meet the training conditions for the reward model, the reward model training and evaluation sub-agent trains or updates the reward model based on the structured feedback data. When the updated reward model's evaluation metrics on the validation or test set meet the corresponding quality admission criteria, the reward model is written into the reward model registry and used for the next round of reward function assembly and reinforcement learning optimization. When the updated reward model does not meet the quality admission criteria, the evaluator prevents the reward model from entering the reward function assembly process, and the orchestrator modifies the workflow according to the corresponding failure state, causing the reward function assembly sub-agent to construct alternative reward components using structure prediction scores, molecular docking scores, or sequence distances or distribution distances between candidate sequences and high-quality experimental samples. Thus, experimental measurement results are not saved separately as termination validation results, but rather as input for the next round of protein sequence design, enabling the system to form a closed loop where computational design, experimental preparation, property detection, feedback learning, and redesign are interconnected.

[0028] This invention also provides a protein sequence computation and experimental closed-loop evolutionary design method based on the above system, comprising: receiving a protein design task and corresponding input data; generating a workflow containing multiple categorized task items by an orchestrator; reviewing the workflow, structured input, task dependencies, and registered products by an evaluator; scheduling multiple specialized sub-agents to complete data processing, generator task adaptation, reward model training and evaluation, reward function assembly, reinforcement learning evolution, and candidate evaluation when the review is passed; executing corresponding model training, candidate sequence generation, structure prediction, molecular docking, and candidate evaluation tasks by a cloud computing and supercomputing platform; and outputting the experimental candidate list and experimental task information to the biological experiment via a candidate sequence output interface. The verification laboratory subsystem includes: a candidate protein preparation unit that prepares candidate protein experimental samples based on candidate protein sequences; a property detection unit that obtains experimental measurement results of candidate protein experimental samples according to experimental task information; an experimental feedback interface that associates experimental measurement results with candidate protein sequences, generation rounds, calculation evaluation results, and experimental samples based on candidate numbers, and converts them into structured feedback data; updating data semantic information, reward models, reward functions, candidate evaluation conditions, memory records, or subsequent workflows based on structured feedback data; generating the next round of candidate protein sequences based on the updated calculation design conditions, and determining whether to continue executing the next round of calculation and experimental loop or output the final candidate protein based on task termination conditions.

[0029] In a preferred embodiment, each entry in the generator product registry includes a source sub-agent, output directory, model configuration file, vocabulary file, initialization model parameters, post-trained model parameters, training input template, chain format, sequence length statistics, and maximum sequence length information. Before running, the reinforcement learning sub-agent selects generator products whose registration status meets the calling conditions from the generator product registry based on the task type, chain format, and sequence length requirements. In a preferred embodiment, each entry in the reward model registry includes reward model parameters, task type, chain type, target label, validation set or test set evaluation metrics, whether the quality admission conditions are met, and source output directory. The reward function assembly sub-agent only uses reward models that meet the quality admission conditions. In a preferred embodiment, each entry in the reward function registry includes a reward function construction program, reward function parameters, reward component weights, output directory, names of reward components that have passed the test, and a test report. The reinforcement learning sub-agent only uses registered and tested reward functions.

[0030] In a preferred embodiment, the system includes a manual review mechanism for at least the following: training configuration review, scientific parameter confirmation, reward intent review, reward weight review, restricted program execution review, and experimental candidate confirmation. For at least one scientific parameter among antigen sequence, fixed pairing sequence, candidate length range, antibody chain format, reward model source, whether experimental feedback is used, whether therapeutic antibody profiling is used, experimental testing items, and the number of experimental candidates, the system executes the corresponding training, reinforcement learning, or experimental task after obtaining structured confirmation information. This manual review mechanism does not replace the automatic execution of the orchestrator, evaluator, and dedicated sub-agents; rather, it serves to establish traceable confirmation nodes before high-computational training, reward function generation, restricted code execution, and biological experiments.

[0031] In a preferred embodiment, the candidate evaluation results output by the evaluator are not only used for the final sorting of candidate protein sequences, but also stored as feedback events in the memory system. The feedback events include at least one of the following: task type, candidate number, candidate sequence, generation round, reward component, evaluation metric, filtering status, and experimental verification priority. In subsequent tasks, the orchestrator or evaluator reads the feedback events through the memory retrieval interface and adjusts the workflow order, reward component selection, candidate filtering threshold, or experimental verification priority based on the feedback events.

[0032] In a preferred embodiment, the biological validation laboratory subsystem writes at least one of the following results as negative feedback into the structured feedback data: expression failure, purification failure, negative binding, insufficient affinity, insufficient stability, excessive aggregation ratio, insufficient functional activity, or failure to pass the exploitability index. It also associates the experimental failure reasons with the corresponding candidate number, candidate protein sequence, generation round, and experimental batch. The reward function assembling agent increases or adjusts expression-related rewards, stability-related rewards, aggregation risk-related rewards, binding-related rewards, or exploitability rewards based on the negative feedback. The memory system stores the experimental failure reasons so that the orchestrator can adjust data processing steps, reward component combinations, candidate filtering conditions, or experimental validation priorities in subsequent similar tasks.

[0033] Compared with existing technologies, this invention has the following beneficial effects: It provides unified heterogeneous computing power, task scheduling, and data storage support for model training, reinforcement learning, structure prediction, molecular docking, and large-scale candidate evaluation through cloud computing and supercomputing platforms; through the collaborative work of orchestrators, evaluators, and multiple specialized sub-agents, it organizes the data processing, model training, reward construction, sequence optimization, and candidate evaluation processes, which originally relied on manual intervention, into a typified workflow that can be reviewed, corrected, and replanned; through product registration and quality admission mechanisms, it reduces the risk of incomplete, untested, or insufficiently quality generators, reward models, and reward functions being called by downstream tasks; and through reward normalization within candidate groups, secondary normalization of training batches, reference model distribution constraints, and candidate... Group-level exclusionary divergence adjustment reduces the risk of candidate sequences concentrating on a few similar patterns during reinforcement learning; a memory system saves workflows, evaluation results, error fields, and experimental feedback, enabling historical execution records to be used for subsequent task planning; a biological validation laboratory subsystem transforms computationally generated candidate sequences into actual protein experimental samples and obtains corresponding experimental measurement results; an experimental feedback interface establishes a traceable relationship between computational candidates, experimental samples, and experimental results, and uses real experimental results for the next round of reward model training, reward function assembly, and candidate sequence optimization; thus forming a unified computational and experimental closed-loop evolutionary design system that covers data-driven protein redesign, de novo binding protein design, de novo antibody design, and experimental data-guided optimization. Attached Figure Description

[0034] Figure 1This is a schematic diagram of the overall architecture of the protein sequence closed-loop evolution design based on a self-evolving multi-agent system of the present invention. The cloud computing and supercomputing platform interacts with the self-evolving multi-agent system through I / O interfaces and utilizes communication, storage, and computing units to provide communication, data storage, and computing resource support for system operation. The task side illustrates protein engineering tasks such as protein redesign, de novo design of binding proteins, and de novo design of antibodies. The self-evolving multi-agent system includes a control layer composed of an orchestrator and an evaluator, and dedicated modules for data processing, task construction, evaluation, task adaptive pre-training, supervised fine-tuning, reward function assembly, reward function model training, and reinforcement learning. The sub-agent layer, composed of sub-agents, has an orchestrator and evaluator working together to complete task planning, execution review, and feedback adjustments, and scheduling each dedicated sub-agent to generate and optimize candidate protein sequences as needed. The results panel displays the design results of candidate proteins, binding proteins, and antibodies output by the system. The biological validation laboratory sequentially performs amino acid sequence synthesis, protein expression vector construction, protein synthesis, and downstream biological experiments on the candidate sequences, and obtains experimental evaluation results such as expression level, affinity, catalytic activity, isoelectric point, and water solubility. The experimental evaluation results are fed back to the self-evolving multi-agent system, driving the candidate sequences to undergo multiple rounds of iterative optimization, ultimately obtaining engineered proteins that meet the design goals.

[0035] Figure 2This is a schematic diagram illustrating the internal composition and workflow of the self-evolving multi-agent system of the present invention. User queries, feedback information, and task data serve as system inputs, which are received by the orchestrator to generate corresponding workflows. The evaluator reviews the workflows, intermediate products, and execution results, and collaborates with the orchestrator to complete task correction and replanning. The task understanding module, through task construction and data processing sub-agents, completes task objective parsing, data standardization, and task semantic confirmation. The model evolution module includes task adaptive pre-training sub-agents, supervised fine-tuning sub-agents, reward model training sub-agents, reward function assembly sub-agents, reinforcement learning sub-agents, and evaluation sub-agents, used to complete task adaptation of the basic model, reward signal construction, reinforcement learning evolution, and candidate result evaluation. Model data... The production module stores and calls the basic generation model, the evolved generation model, and the reward model, and receives the model products generated during the model evolution process. The execution sandbox isolates and executes the programs, models, and tools generated or called by each sub-agent, and feeds back the execution results to the corresponding functional modules. The memory system includes short-term memory, long-term memory, a memory compressor, an insight library, and filters. The short-term memory is truncated and compressed when the context window meets preset conditions. The extracted valid insights are filtered and written into the long-term memory. The orchestrator and evaluator call historical task information, execution experience, and optimization strategies through memory retrieval tools. After task understanding, model evolution, execution review, and memory update, the system outputs the task execution results and continuously updates the workflow and model assets using the results and historical feedback.

[0036] Figure 3 This diagram illustrates the multi-objective group reward decoupling normalization algorithm used by the reinforcement learning sub-agent in this invention. The policy model generates a set of candidate outputs based on the input task and obtains the reward value of each candidate output under different optimization objectives through multiple reward functions. For each reward objective, the reward value is first independently normalized within the candidate group to obtain the corresponding objective advantage. Subsequently, the multiple objective advantages are weighted and combined according to the preset weights of each reward objective to obtain the total advantage of each candidate output. This total advantage is then normalized a second time within the candidate group to form the final advantage used to update the policy model. During training, the distribution distance between the policy model and the reference model is constrained by KL divergence to limit the degree of deviation of the policy model relative to the reference model. Simultaneously, a characterization of KL divergence is introduced, and an exclusionary regularization term is constructed based on the JS divergence between candidate outputs to increase the distribution difference between candidate outputs, maintain the diversity of candidate sequences, and reduce the risk of pattern collapse, thereby achieving a balance between multi-objective reward optimization, policy stability, and effective exploration.

[0037] Figure 4 The results show a comparison of different reinforcement learning algorithms in terms of reward optimization ability and candidate sequence diversity; among them, Figure 4A shows the reward value changes of ProteinRL, DyNA-PPO, PPO, GRPO, DAPO, GSPO and GDPO during the training iteration process. PPO, GRPO, DAPO, GDPO and GSPO can make the reward value converge quickly, ProteinRL does not achieve effective optimization, and DyNA-PPO exhibits periodic oscillations. Figure 4 B shows the policy entropy changes of different algorithms during the training process. The policy entropy of PPO, GRPO, DAPO, GDPO and GSPO gradually decreases and tends to converge. The policy entropy of ProteinRL does not converge significantly. The policy entropy of DyNA-PPO fluctuates periodically. Figure 4 C shows the distribution of reward model prediction scores for the training set and sequences generated by different algorithms. The prediction scores of the sequences generated by each reinforcement learning algorithm are improved relative to the training set, with GDPO having a higher average reward model prediction score. Figure 4 D shows the hydrophobicity distribution of the training set and sequences generated by different algorithms. The hydrophobicity of the DyNA-PPO generated sequences is significantly shifted relative to the training set distribution. Figure 4 E shows the repetition rate of sequences generated by different algorithms. PPO, GRPO, DAPO, GDPO, and GSPO generate sequences with a high repetition rate. Figure 4 F shows the edit distance between sequences generated by different algorithms.

[0038] Figure 5 This is a schematic diagram illustrating the reinforcement learning optimization results of GDPO and its variants; where, Figure 5 A shows the changes in reward values ​​for GDPO, GDPO+D, DMA, DynClip, and RepKL during reinforcement learning training; Figure 5 B shows the average word-level entropy change of each algorithm during the training process; Figure 5 C shows the reward model score distribution of the sequences generated by each algorithm; Figure 5 D shows the reward model scores of the top 10 high-scoring sequences among the non-repeating sequences generated by each algorithm; Figure 5 E shows the distribution of the non-repeating sequences generated by each algorithm after one-hot encoding and UMAP dimensionality reduction; Figure 5 F shows the GRAVY hydrophobicity distribution of the training set and the sequences generated by each algorithm; Figure 5 G shows the internal pairwise edit distances between non-repeating sequences generated by each algorithm, normalized to the sequence length. Figure 5 H shows the sequence repetition rate for each algorithm in three independent generation experiments.

[0039] Figure 6 This is a dynamic illustration of sequence space exploration and discovery during the reinforcement learning training process of GDPO+RepKL and GDPO; among them, Figure 6 A and Figure 6 B shows the minimum spanning tree topology constructed by the non-repeating sequences generated by GDPO+RepKL and GDPO throughout the training process. Each node represents a non-repeating sequence, the node color represents the number of training iterations corresponding to the generation of the sequence, red nodes represent the sequence generated in the last iteration, and the lines between nodes represent the sequence connection relationship in the minimum spanning tree. Figure 6 C shows the total weight of the minimum spanning tree corresponding to GDPO+RepKL and GDPO, where the total weight represents the cumulative value of the distance traversed by the minimum spanning tree in the sequence space; Figure 6 D shows the diameter of the minimum spanning tree corresponding to GDPO+RepKL and GDPO, where the diameter represents the maximum path length between any two nodes in the corresponding minimum spanning tree.

[0040] Figure 7 This diagram illustrates the application and performance evaluation of the self-evolving multi-agent system of the present invention in data-driven protein redesign tasks targeting different structural types; wherein, Figure 7 A illustrates the process by which the system, based on the target protein redesign task and corresponding deep mutation scanning data, forms a task-specific protein design model and generates candidate sequences through iterative reasoning. Figure 7 B illustrates the optimization process, including data processing, task-adaptive pre-training, reward model training, reward function assembly, reinforcement learning, and evaluation, as well as the optimization trajectory of candidate sequences in the target property space. Figure 7 C shows the ESM2 log-likelihood evaluation results of SEMAS, ProteinMPNN, ESM-IF1, and AFDesign for the AAV2, HIS7, Kir2.1, and EcCM design sequences. The names of the structural information driven models are marked in magenta. Figure 7 D shows the training set, the top 32 sequences in the training set, and the functional scores of the SEMAS designed sequences on four targets, where AAV2 and HIS7 correspond to production fitness, Kir2.1 corresponds to membrane surface expression level, and EcCM corresponds to catalytic activity. Figure 7 E shows a heatmap of the prediction template modeling score pTM for different method design sequences, where the pTM is used to characterize the global structure confidence. Figure 7 F shows a heatmap of the predicted local distance difference test scores (pLDDT) for sequences designed by different methods. The pLDDT is used to characterize residue-by-residue structural confidence. Figure 7 G shows the superposition results of the protein structures designed by SEMAS for AAV2, HIS7, Kir2.1 and EcCM with the corresponding wild-type reference structures, and indicates the edit distance of the designed sequence relative to the wild-type sequence and the root mean square deviation (RMSD) of the superposition. Figure 7 C and Figure 7 The significance of D was evaluated using a two-sided Mann-Whitney U test corrected for Bonferroni, with * indicating P < 0.05, ** indicating P < 0.01, and *** indicating P < 0.001.

[0041] Figure 8 A schematic diagram showing the secondary structure composition of sequences obtained from different protein design methods and the relationship between predicted water solubility and target-specific functional indicators; wherein, Figure 8 A shows the proportions of spiral, folded, and coiled structures of wild-type sequences and SEMAS, ProteinMPNN, ESM-IF1, and AFDesign for AAV2, HIS7, Kir2.1, and EcCM designs; Figure 8 B to Figure 8 E shows the relationship between the predicted water solubility and the corresponding functional indicators for the AAV2, HIS7, Kir2.1 and EcCM design sequences using different methods. The predicted water solubility is represented by the negative average hydrophobicity index (-GRAVY), and a higher -GRAVY value indicates a higher predicted water solubility. Figure 8 B and Figure 8 The vertical axis of C represents the production adaptability of AAV2 and HIS7, respectively. Figure 8 The ordinate of D represents the expression level of Kir2.1 on the membrane surface. Figure 8 The ordinate of E represents the catalytic activity of EcCM; each data point represents the average value of the corresponding group, the error bar represents the standard deviation, and the name of the structure information driven model is indicated in magenta.

[0042] Figure 9 This is a schematic diagram illustrating the application and performance evaluation of the self-evolving multi-agent system of the present invention in de novo protein design and self-evolution tasks; wherein, Figure 9 A illustrates the process by which the system performs cold-start design of binding proteins based on target protein information, and iteratively optimizes the binding proteins through data processing, reward function assembly, reinforcement learning, and evaluation sub-agents based on human feedback. Figure 9 B to Figure 9 D shows the evaluation results of SEMAS, supervised fine-tuning model SFT, RFdiffusion, and AFDesign for designing binding proteins for the three targets CD47, IL-7Ra, and PD-L1. The names of the structure information driven methods are indicated in magenta. Figure 9 B shows the ipTM score used to characterize the overall confidence of the predicted inter-chain interface; Figure 9 C shows the interface-specific confidence level ipSAE calculated based on the binding interface residues; Figure 9 D demonstrates the sequence novelty of the designed binding protein; Figure 9E shows the global structural confidence pTM of the binding protein complexes designed for the three targets using different methods; Figure 9 F shows the residue-by-residue structure confidence level pLDDT for design results from different methods; Figure 9 G shows the structure of a representative CD47-binding protein complex designed by SEMAS and its local binding interface, where CD47 is shown in pink and the binding protein is shown in blue, and the ipTM, ipSAE and interface residues of the complex are labeled. Figure 9 B to Figure 9 The significance of D was evaluated using a two-sided Mann-Whitney U test corrected for Benjamini-Hochberg false discovery rate. * indicates P < 0.05, ** indicates P < 0.01, and *** indicates P < 0.001.

[0043] Figure 10 This diagram illustrates the interface, structure, and sequence parameters of the protein-binding system designed before and after feedback-guided self-evolution in the self-evolving multi-agent system of this invention; wherein, Figure 10 A shows the normalized results of the binding proteins designed in the first and second rounds of SEMAS on five indicators: ipTM, pLDDT, pTM, ipSAE, and sequence novelty. The percentages represent the change ratio of the second round to the first round. Figure 10 B and Figure 10 C shows the representative IL-7Ra binding protein complex and PD-L1 binding protein complex designed in the second round of SEMAS, respectively. The target protein is shown in pink and the designed binding protein is shown in blue. The magnified view shows the binding interface and corresponding contact residues, and the ipTM and ipSAE scores of each complex are marked. Figure 10 D to Figure 10 F shows the relationship between sequence novelty of the designed sequences by SEMAS, RFdiffusion, and AFDesign and the residue-by-residue structural confidence of the binding protein, pLDDT. Figure 10 G to Figure 10 I shows the relationship between sequence novelty and global structural confidence (pTM) of the binding protein designed by SEMAS, RFdiffusion, and AFDesign, respectively. Figure 10 D to Figure 10 The dashed line in I represents the linear regression fitting result, r represents the Pearson correlation coefficient, and the name of the structural information driven method is marked in magenta.

[0044] Figure 11 This is a schematic diagram illustrating the application and performance evaluation of the self-evolving multi-agent system of the present invention in the antibody de novo design task; wherein, Figure 11A illustrates the task of de novo antibody design based on the target antigen sequence in the absence of reference antibody data; Figure 11 B illustrates the antibody design process, which includes data processing, reward function assembly, reinforcement learning, and evaluation. Figure 11 C to Figure 11 F shows the predicted structural quality and binding confidence of antibodies designed by SEMAS, MAGE, DiffAb, and RFAntibody against seven antigen targets: Integrin-α-1, HEWL, HER2, VEGF-A, H1N1, Ang2, and Spike HR2. The names of the structural information-driven methods are indicated in magenta. Figure 11 C shows the ipTM score used to characterize the overall confidence level between chains; Figure 11 D shows the ipSAE score used to characterize the confidence of the combined interface; Figure 11 E shows the pLDDT score used to characterize residue-by-residue structural confidence; Figure 11 F shows the pTM score used to characterize the global folding confidence; each bar represents the mean, the error bar represents the standard deviation, and the scatter plot represents a single antibody design. Each method generates 32 antibody designs for each target. Significance was evaluated using a two-sided Mann-Whitney U test corrected for Benjamini-Hochberg false discovery rate. * indicates P < 0.05, ** indicates P < 0.01, and *** indicates P < 0.001.

[0045] Figure 12 This diagram illustrates the evaluation of the individual folding stability and VH-VL chain pairing quality of antibodies designed using different methods using AlphaFold-Multimer; where, Figure 12 A to Figure 12 E uses a radar chart to show the average values ​​of various indicators for antibodies designed by SEMAS, MAGE, DiffAb, and RFAntibody against seven antigen targets: Integrin-α-1, HEWL, HER2, VEGF-A, H1N1, Ang2, and Spike HR2. Each axis represents an antigen target, and the radial distance represents the average value of the corresponding method on that target. The names of the structure information-driven methods are indicated in magenta. Figure 12 A shows ipTM used to characterize the confidence level of pairing between the antibody heavy chain variable region and the light chain variable region; Figure 12 B shows the ipSAE used to characterize the quality of the VH-VL pairing interface; Figure 12 C shows the pTM used to characterize the overall structural quality of the antibody; Figure 12 D shows the pLDDT used to characterize the residue-by-residue structural confidence of the antibody; Figure 12E shows the ipLDDT used to characterize residue-by-residue confidence at the VH-VL pairing interface.

[0046] Figure 13 A schematic diagram illustrating the Pareto front analysis of different antibody design methods for achieving both structural confidence and sequence novelty objectives; among which... Figure 13 A shows the relationship between the average inter-chain confidence ipTM and the average sequence novelty obtained by SEMAS, MAGE, DiffAb and RFAntibody for each antigen target. Each data point represents a combination of a design method and an antigen target. Figure 13 B shows the Pareto front hypervolume indices for each method in the ipTM-sequence novelty space, calculated using a common reference point. Figure 13 C shows the number of times each method enters the Pareto front in the ipTM-sequence novelty space, and the percentage above the bar represents the proportion of the corresponding method to the total number of Pareto front occurrences; Figure 13 D shows the relationship between the average binding interface confidence (ipSAE) and average sequence novelty obtained by each method for different antigen targets. Each data point represents a combination of a design method and an antigen target. Figure 13 E shows the Pareto front hypervolume indices for each method in the ipSAE-sequence novelty space, calculated using a common reference point. Figure 13 F shows the number of times each method enters the Pareto front in the ipSAE sequence novelty space and the proportion thereof; the names of structural information-driven methods are indicated in magenta.

[0047] Figure 14 This is a schematic diagram illustrating the application and evaluation of the self-evolving multi-agent system of the present invention in a wet experimental data-guided antibody design task; wherein, Figure 14 A illustrates a two-round antibody design process. The first round uses the antigen sequence as input for cold-start antibody design, and the second round introduces processed deep mutation scanning data to optimize the antibody through data guidance. Figure 14 B illustrates the data-driven optimization process, which includes data processing, task-adaptive pre-training, reward model training, reward function assembly, reinforcement learning, and evaluation. Figure 14 C shows the training set sequences, the top 32 sequences screened by wet experiments, and the predicted affinity scores of SEMAS (+Data) designed antibodies against five antigens: HEWL, VEGF-A, H1N1, Ang2, and Spike HR2. The bars represent the mean, the error bars represent the standard deviation, and the scatter plots represent individual sequences. Figure 14D shows the therapeutic antibody profiles of cold-start SEMAS-designed antibodies, Top32 sequences, and SEMAS(+Data)-designed antibodies. TAP is represented by the ratio, bars represent the average value of each target, error bars represent the standard deviation, and connected data points represent the corresponding values ​​of each target. Figure 14 E to Figure 14 H shows the distribution of TAP exploitability indicators for clinical antibodies, cold-start SEMAS-designed antibodies, Top32 sequences, and SEMAS (+Data)-designed antibodies. The red dashed lines represent the risk thresholds of the corresponding indicators. Figure 14 E indicates the surface hydrophobicity distribution. Figure 14 F indicates the distribution of negative charge on the surface. Figure 14 G indicates the surface positive charge distribution. Figure 14 H indicates the distribution of charge symmetry parameters; Figure 14 The significance of C was evaluated using a two-sided Mann-Whitney U test corrected for Bonferroni. Figure 14 The number of pass and fail groups in D was evaluated using a Bonferroni-corrected chi-square test, where * indicates P < 0.05, ** indicates P < 0.01, and *** indicates P < 0.001.

[0048] Figure 15 This is a schematic diagram showing the complementarity-determining region lengths and amino acid composition of the cold-start SEMAS, Top32, and SEMAS(+Data) antibody sequences; among them, Figure 15 A shows the total length distribution of complementary determinant regions for cold-start SEMAS-designed antibodies, Top32 sequences, SEMAS (+Data)-designed antibodies, and clinical antibodies. The red dashed line represents the exploitability risk threshold defined by the Therapeutic Antibody Profile (TAP), and the percentage on the right represents the proportion of sequences in each group that fall within the corresponding threshold range. Figure 15 B shows the overall amino acid frequency distribution of the three antibody sequences: cold-start SEMAS, Top32, and SEMAS(+Data). Figure 15 C to Figure 15 H shows the amino acid composition distribution of the three antibody sequences in the heavy chain complementarity-determining regions CDRH1, CDRH2, and CDRH3, and the light chain complementarity-determining regions CDRL1, CDRL2, and CDRL3, respectively, using radar charts.

[0049] Figure 16 A schematic diagram illustrating the alternative optimization strategies and structural evaluations used in the design of antibodies for HER2 and Integrin-α-1; among which, Figure 16 A illustrates the data-guided optimization process for the HER2 antibody. In cases where the reward model fails to meet quality standards, the system employs a proxy reward function based on molecular docking scores to continue executing reward function assembly, reinforcement learning, and evaluation. Figure 16 B uses a radar chart to illustrate the cold-start SEMAS-designed antibody, the Top32 wet experimental screening sequences, and the ipTM, ipSAE, pTM, and pLDDT / 100 indicators of the SEMAS (+Data)-designed antibody against HER2. Figure 16 C shows the Top32 sequence and the representative antibody-antigen complex structure formed by the SEMAS (+Data) designed antibody and HER2. The residues in the antibody VH-VL domain that interact with the antigen are marked in red, and the corresponding model edit distances ED, ipSAE and ipTM are also indicated. Figure 16 D illustrates a simplified data-guided optimization process for the Integrin-α-1 antibody. When the dataset size is limited, the system employs a data processing, reward function assembly, reinforcement learning, and evaluation process that includes a distributed traction reward function. Figure 16 E uses a radar chart to illustrate the cold-start SEMAS-designed antibody, the Top32 wet experimental screening sequences, and the ipTM, ipSAE, pTM, and pLDDT / 100 metrics of the SEMAS (+Data)-designed antibody against Integrin-α-1. Figure 16 F shows the representative antibody-antigen complex structures formed by the Top32 sequence and SEMAS (+Data) designed antibodies with Integrin-α-1. The residues in the antibody VH-VL domain that interact with the antigen are marked in red, and the edit distances ED, ipSAE, and ipTM of each model are also indicated.

[0050] Figure 17 A comparative diagram showing the time required to generate a single sequence for different protein design methods; where, Figure 17 A shows the average time for SEMAS, AFDesign, ESM-IF1, and ProteinMPNN to generate a single sequence in a data-driven protein redesign task; Figure 17 B shows the average time for SEMAS, AFDesign, and RFdiffusion to generate a single sequence in a de novo protein design task; Figure 17 Figure C shows the average time to generate a single sequence using SEMAS, MAGE, RFAntibody, and DiffAb in antibody design tasks; the vertical axis of each figure represents the number of seconds required to generate a single sequence on a logarithmic scale, the values ​​above the bars represent the average generation time, the multiples represent the multiples of the average generation time of the corresponding method relative to SEMAS, the scatter plots represent the results of each independent test, and the names of the structure information driven methods are marked in magenta.

[0051] Detailed description Terminology Definition Unless otherwise expressly defined, the terms and technical terms used in this specification shall be understood to have the meanings commonly given to them by those skilled in the art. Methods, materials, and techniques related to protein design, computational biology, artificial intelligence, multi-agent systems, machine learning, molecular biology, protein engineering, and experimental validation can be implemented using equivalent methods known or available in the art. When dealing with operations such as sequence analysis, model training, structure prediction, candidate screening, expression purification, and activity evaluation, reference can be made to the relevant manufacturer's instructions, conventional practices in the art, prevailing technical specifications, or the embodiments described herein. Publications, patents, and other references cited herein are incorporated herein by reference in their entirety; however, such references are for illustrative purposes only and do not constitute an endorsement that any content therein is common knowledge in the art.

[0052] intelligent agent An intelligent agent refers to a functional unit capable of performing task decomposition, information reasoning, tool invocation, result evaluation, or decision output around a predetermined goal. Intelligent agents can be implemented based on rules, models, or a combination of rules and models; their key lies in their ability to receive input, process context, and output structured results that can be used by subsequent modules. In this invention, an intelligent agent does not refer to a subject with anthropomorphic consciousness, but rather to a computational module with task autonomy capabilities.

[0053] Self-evolving multi-agent systems (SEMAS) SEMAS refers to a closed-loop collaborative system consisting of an orchestrator, an evaluator, multiple specialized sub-agents, and a memory system. In protein design tasks, this system continuously generates, reviews, corrects, and optimizes workflows based on task input, historical execution results, and external feedback, enabling candidate protein sequences to gradually converge towards the design target through multiple iterations. The core of SEMAS lies not in a single generation of results, but in the cyclical evolutionary mechanism of "generation-evaluation-memory-replanning".

[0054] Specialized sub-smart agents These are intelligent agents designed for specific task stages, with clearly defined responsibilities and functions focused on a particular type of operation in the protein design process, such as data processing, generator task adaptation, reward model training and evaluation, reward function assembly, reinforcement learning evolution, structural evaluation, or exploitability evaluation. These specialized sub-agents can work in parallel, sequentially, or conditionally to coordinate and support the overall workflow execution.

[0055] Arranger This refers to the control module used to receive protein design tasks, break down task objectives, generate workflows, allocate subtasks, and schedule dedicated sub-agents to execute them. The orchestrator's role is not merely to "issue commands," but to dynamically determine subsequent processes based on task stages, historical memories, and evaluation feedback, and can also reorganize the task structure when necessary.

[0056] evaluator This refers to the review module used to perform rule review, semantic review, and quality evaluation on workflows, intermediate products, candidate sequences, and task execution results. The evaluator not only judges the quality of the results, but also whether the task path is correct, whether the input and output are consistent, and whether the current execution should continue, be corrected, or replanned.

[0057] Workflow Workflow refers to the ordered execution path generated by the orchestrator to complete the protein design task. Workflow can contain multiple steps, multiple subtasks and their dependencies, and can also be incrementally modified, inserted, deleted or rearranged based on task feedback.

[0058] Tool call This refers to the behavior of an orchestrator or sub-agent calling external functions, models, algorithms, evaluators, databases, or data processing modules as needed for a task. Tool calls are the foundation for achieving task autonomy and module collaboration, and are not limited to a single program interface.

[0059] Large Language Model This refers to a general-purpose foundational model pre-trained on large-scale text or multimodal corpora, capable of performing natural language understanding, task planning, inference generation, tool selection, or structured information output under contextual constraints. In this invention, the large language model can serve as the core inference engine for an arranger, evaluator, or other specialized sub-agents, or as the underlying capability source for a generator or review module.

[0060] generator This refers to the model, algorithm, or module used to generate candidate protein sequences. The generator can be a large language model, sequence model, diffusion model, evolutionary search model, or a combination thereof, and its output is the candidate sequences that enter the subsequent evaluation and optimization process.

[0061] Large language model-driven generator This refers to a generation module that uses a large language model for protein sequence generation, task planning, or candidate strategy output. Its inputs may include task descriptions, constraints, historical sequences, evaluation results, or memory information, and its outputs may include sequence drafts, design suggestions, candidate pools, or next action plans.

[0062] Supervision and fine-tuning Supervised fine-tuning refers to the process of further training a pre-trained model using labeled data with clear input-output correspondences to make the model more suitable for a specific task. In this invention, supervised fine-tuning can be used to adapt the generator to protein redesign, binding protein design, or antibody design tasks, making the model output more consistent with sequence constraints, structural requirements, and evaluation objectives.

[0063] Task Adaptive Pre-training Task-adaptive pre-training refers to further training the model on top of general pre-training, using corpora, sequences, structural information, or evaluation information designed for a specific protein task. This aims to improve the model's understanding and adaptation to the task's distribution. Compared to supervised fine-tuning, task-adaptive pre-training focuses more on allowing the model to absorb domain statistical features and task priors.

[0064] Reward Model A reward model is a model used to output evaluation signals based on the properties, structure, binding ability, stability, exploitability, or other task objectives of candidate sequences. Reward models can score a single metric or jointly model multiple metrics, and their output can serve as the basis for subsequent reinforcement learning or selection steps.

[0065] reward function The reward function is an optimization objective function formed by combining the output of the reward model, structural evaluation results, exploitability indicators, task constraints, and design goals. The reward function guides the generator or reinforcement learning module in iterative optimization and can dynamically adjust its weights according to different task stages.

[0066] Collaborative training of supervised fine-tuning and reward optimization This refers to a training method that first uses supervised fine-tuning to enable the model to acquire basic task capabilities, and then further optimizes the model's behavior through reward-based models or reinforcement learning. This approach can balance the basic controllability of the model with goal-orientedness, and is suitable for complex protein design scenarios.

[0067] reinforcement learning This refers to the learning process of iteratively optimizing the generator strategy based on the reward function. In this invention, reinforcement learning is used to drive candidate sequences toward higher goal consistency, better structural rationality, or better exploitability.

[0068] Memory system This refers to the module used to store task status, historical workflows, candidate sequences, evaluation results, failure reasons, correction information, and design experience. The memory system not only performs storage functions but also provides a callable historical context for subsequent task orchestration, evaluation, and replanning.

[0069] Reflection-Evolutionary Cycle This refers to a closed-loop process by which the system replans and re-optimizes the workflow, reward strategy, candidate sequences, or evaluation logic based on the design results of the current round, evaluation feedback, and memory information. This loop is the core mechanism that distinguishes SEMAS from single-round generative methods.

[0070] Candidate sequences Candidate sequences refer to protein sequences generated by a generator or sub-agent and subsequently evaluated, screened, or optimized. Candidate sequences can be complete proteins, protein fragments, binding proteins, antibody sequences, or variants thereof.

[0071] Candidate Group A candidate group is a collection of multiple candidate protein sequences generated by a protein sequence generator within the same sampling or optimization round, based on the same protein design task input or conditional prompts. The candidate protein sequences in a candidate group share the same task objective and input conditions, but may have different amino acid compositions due to differences in sampling processes, generation paths, or strategy probability distributions. A candidate group is the basic data unit for intra-group reward normalization, candidate sequence comparison, diversity evaluation, and relative strategy optimization. Unless otherwise specified, candidate protein sequences within a candidate group are not required to have exactly the same length.

[0072] Lexicon This refers to the basic sequence unit used by the protein sequence generator when performing sequence encoding, autoregressive generation, or probability calculation. The term can represent a single standard amino acid, a non-standard amino acid, an inter-strand separator, a sequence start symbol, a sequence stop symbol, a filler symbol, or other preset sequence symbols. In this invention, the protein sequence generator outputs the conditional probability of each candidate term in a preset vocabulary appearing at the next generation position based on the already generated term sequence, and selects or samples the next term accordingly.

[0073] Valid lexical elements Valid terms refer to the terms in a candidate protein sequence that participate in reward calculation, policy loss calculation, or divergence calculation. Valid terms typically include terms representing amino acids and may include sequence start symbols, sequence end symbols, or inter-strand separators depending on the model configuration; padding symbols and other terms excluded by the valid position mask are not valid terms in the corresponding calculations. For candidate protein sequences of different lengths, the positions of valid terms in each sequence can be determined using the valid position mask, and the difference in policy probability distribution is calculated only at the corresponding generation positions where both candidate sequences have valid terms.

[0074] Current strategy model This refers to a protein sequence generation model that accepts parameter updates during the current reinforcement learning optimization phase and is used to generate candidate protein sequences and output conditional probability distributions of terms. The current policy model, based on the protein design task input and the generated sequence prefixes, outputs the conditional probability of each term in the predefined vocabulary appearing at the next generation position. The parameters of the current policy model are updated based on the pruning agent policy loss, distribution constraints relative to the reference model, and diversity adjustment signals within the candidate group.

[0075] Old strategy model The old policy model refers to the policy model used as the benchmark for calculating the policy probability ratio during the current policy update or the current stage of the policy update process. The old policy model can be copied or fixed from the current policy model before performing the corresponding parameter update, and is used to calculate the ratio between the output probability of the current policy model and the output probability of the policy model before the update, under the same task input and the same generated lexical units. By using the old policy model as a relatively stable probability benchmark, the magnitude of a single policy update can be limited, reducing drastic policy changes during reinforcement learning training.

[0076] Reference Model The reference model is used to constrain the degree to which the current policy model deviates from a preset protein sequence distribution. The reference model can be obtained from a basic protein sequence generation model that has not been optimized by reinforcement learning, a model pre-trained adaptively for the task, or a model fine-tuned under supervision. During the corresponding reinforcement learning optimization process, the parameters of the reference model are not updated based on the current joint loss. The difference in probability distribution between the current policy model and the reference model can be characterized by Kullback-Leibler divergence to reduce the risk of the current policy model deviating from a legitimate, natural, or structurally sound protein sequence distribution due to excessive pursuit of rewards.

[0077] Rewards This refers to a numerical evaluation signal calculated for a single design objective, constraint, or evaluation dimension of a candidate protein sequence. The reward components can originate from reward models, sequence analysis methods, structure prediction models, molecular docking procedures, empirical rules, or experimental data, and may include at least one of the following: functional prediction rewards, sequence naturalness rewards, amino acid legitimacy rewards, length constraint rewards, structure confidence rewards, binding interface rewards, diversity rewards, solubility-related rewards, and exploitability rewards. Multiple reward components can be normalized individually and combined according to preset or dynamically determined weights to form a comprehensive evaluation signal for reinforcement learning optimization.

[0078] Advantages Advantage refers to a numerical value used to characterize the relative superiority or inferiority of a candidate protein sequence compared to other candidate protein sequences in the same candidate group or compared to a preset benchmark. Advantage can be obtained by normalizing and combining one or more reward components, and is used to adjust the direction and magnitude of the contribution of each effective term in the candidate protein sequence to the policy update. Positive advantages are typically used to increase the probability that the current policy model will generate the corresponding sequence or path again, while negative advantages are typically used to decrease the corresponding generation probability. Unless otherwise stated, the advantages in this invention can be sequence-level numerical values ​​and are assigned or broadcast to each effective term position of the corresponding candidate protein sequence.

[0079] Within-group normalization and quadratic normalization Intra-group normalization refers to standardizing the same reward component of each candidate protein sequence within the same candidate group based on its mean and standard deviation within that candidate group. This reduces the impact of the numerical scale and distribution range of different reward components on the combination result. The normalized reward components can then be weighted and combined according to their respective weights to form the overall advantage of each candidate protein sequence. Secondary normalization involves standardizing the overall advantage again within a training batch containing one or more candidate groups, based on the mean and standard deviation of the overall advantage of each candidate protein sequence in that batch, to obtain the final advantage used for policy loss calculation. To avoid numerical instability caused by a standard deviation of zero or close to zero, a preset stability constant can be introduced into the normalization calculation.

[0080] Trimming agent strategy loss This refers to the policy optimization loss, constructed from the ratio of the conditional probabilities of generating corresponding effective lexical units based on the current policy model and the old policy model, the advantage corresponding to the candidate protein sequence, and a preset pruning threshold. The pruning process is used to limit the policy probability ratio within a preset range, thereby preventing excessive probability changes in the current policy model relative to the old policy model during a single update. For the same effective lexical unit, the product of the unpruned policy probability ratio and the advantage, and the product of the pruned policy probability ratio and the advantage, can be calculated separately, and the corresponding loss can be determined based on these two calculations. The pruned proxy policy loss is used to maintain the stability of the training process while improving the target reward.

[0081] Kullback-Leibler divergence (KL divergence) KL divergence is an asymmetric measure of the difference in probability distributions, used to characterize the degree of deviation of one probability distribution from another. In this invention, KL divergence is primarily used to characterize the degree of deviation of the term conditional probability distribution output by the current policy model from the corresponding term conditional probability distribution output by the reference model, and serves as a constraint term in policy updates. KL divergence can be calculated at the term level, and the sequence-level KL divergence is obtained by averaging across effective generation positions based on the effective position mask of the candidate protein sequence. By introducing KL divergence constraints into the joint loss to be optimized, the current policy model can be limited from deviating excessively from the prior distribution of the protein sequence represented by the reference model due to reward optimization.

[0082] Jensen-Shannon divergence (JS divergence) JS divergence is a symmetric distribution difference measure determined based on the KL divergence of two probability distributions relative to their mixed probability distribution. In this invention, JS divergence is used to compare the lexical conditional probability distributions of two candidate protein sequences within the same candidate group at corresponding generation positions. For two candidate protein sequences, the lexical conditional probability distributions output by the current strategy model under their respective sequence prefix conditions on a preset vocabulary can be obtained, and the lexical-level JS divergence is calculated based on the equally weighted mixed probability distribution of the two distributions. The larger the JS divergence, the greater the difference in the subsequent lexical probability distributions given by the current strategy model at corresponding positions in the two generation paths.

[0083] Candidate group-level exclusionary JS divergence This refers to the group-level distributional difference obtained by aggregating the term-level JS divergence among multiple candidate protein sequences within the same candidate group, and used for policy optimization in a way that increases the difference in probability distribution between candidate generation paths. For any two candidate protein sequences within a candidate group, the term-level JS divergence can be calculated first at the common effective generation positions of the two sequences, and then the average of all common effective generation positions can be obtained to obtain the sequence pair-level JS divergence; subsequently, the sequence pair-level JS divergence of the disordered and non-repeating candidate sequence pairs within the candidate group can be averaged to obtain the candidate group-level repulsive JS divergence. "Repulsive" does not refer to another independent definition of divergence, but rather to the fact that this JS divergence participates in policy updates in the joint loss in the direction that promotes its increase, thereby reducing the degree to which candidate sequences are concentrated in a few similar generation patterns.

[0084] Pattern collapse Pattern collapse refers to the phenomenon where protein sequence generators, during optimization, gradually concentrate their generation probability on a few high-reward sequences, local sequence patterns, or highly similar generation paths. This leads to an increase in the repetition rate of generated sequences, a decrease in internal edit distance, a reduction in sequence spatial coverage, or a decline in effective exploration capability. Pattern collapse does not necessarily manifest as a decrease in reward value; in some cases, the generator can obtain a high reward model score, but the generated results are concentrated in a narrow high-reward region. This invention reduces the risk of pattern collapse by adjusting the probability distribution differences within candidate groups, while maintaining the optimization of the target properties.

[0085] Quality Access Conditions This refers to the preset evaluation conditions that generator products, reward models, reward functions, or other intermediate products must meet before they can be registered and used by downstream modules. Quality admission conditions may include complete data format, consistency between input / output fields and the calling interface, passing of procedural tests, and one or more evaluation metrics reaching corresponding thresholds on the validation or test set. For regression reward models, the evaluation metrics may include correlation coefficient, ranking correlation coefficient, mean squared error, or mean absolute error; for classification reward models, the evaluation metrics may include accuracy, precision, recall, F1 score, or area under the curve. Products that do not meet the quality admission conditions are not entered into the corresponding product registry or are marked as unusable for subsequent reward function assembly or reinforcement learning tasks.

[0086] Structured feedback data This refers to data formed by converting wet experimental results, deep mutation scan results, or other external experimental information into a standardized format, which is then used for input into the system's evaluation, memory, or reprogramming process.

[0087] External feedback External feedback refers to feedback information originating from outside the system but capable of being imported into the memory system, evaluator, or orchestrator. External feedback does not alter the core closed-loop architecture of SEMAS, but it can enhance the system's experimental coupling capabilities.

[0088] Protein redesign This refers to the targeted optimization, partial replacement, rearrangement, combination, or overall modification of an existing protein sequence, structure, or function to obtain a protein sequence that meets the new objective.

[0089] Protein-binding design This refers to the process of designing protein sequences that can interact specifically or with high affinity with a specific target molecule. The target can be a protein, peptide, or other biomolecule; the design can be de novo or an optimized design based on an existing scaffold.

[0090] Antibody design This refers to the process of designing, optimizing, or redesigning the sequence, structure, or key regions of an antibody or antibody fragment, focusing on objectives such as antigen recognition, binding interface, structural stability, expressibility, or exploitability. The antibody may include full-length antibodies, single-chain antibodies, fragment antibodies, or their functional derivatives. Detailed Implementation

[0091] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the technical features in the following embodiments can be combined with each other.

[0092] Example 1: General Self-Evolving Multi-Agent Process like Figure 1 As shown, this embodiment provides a closed-loop evolutionary design process for protein sequences based on a self-evolving multi-agent system. This process is jointly implemented by a task input terminal, a cloud computing and supercomputing platform, a self-evolving multi-agent system, a candidate result output terminal, and a biological verification laboratory. The protein design task input by the user can be a data-driven protein redesign, a de novo design of binding proteins, a de novo design of antibodies, or a protein and antibody optimization task guided by experimental data. The task input can include at least one of the following: the amino acid sequence of the target protein or antigen, candidate sequences, deep mutation scanning data, functional evaluation tags, binding experimental data, sequence length constraints, fixed pairing sequences, and expected optimization targets. The cloud computing and supercomputing platform includes a communication unit, a storage unit, a computing unit, and an I / O interface. The communication unit is used to transmit task instructions, model parameters, candidate sequences, and evaluation results. The storage unit is used to save training data, model checkpoints, reward models, reward functions, candidate sequences, and historical task records. The computing unit is used to perform task adaptive pre-training, supervised fine-tuning, reward model training, reinforcement learning evolution, structure prediction, and candidate evaluation. The I / O interface is used to realize data interaction between the cloud computing and supercomputing platform and the self-evolving multi-agent system. The self-evolving multi-agent system generates candidate protein, binding protein, or antibody sequences based on task input and outputs candidate sequences that meet preset conditions to a biological validation laboratory. The biological validation laboratory performs amino acid sequence synthesis, protein expression vector construction, protein expression and purification, and downstream biological experiments on the candidate sequences to obtain expression levels, affinity, catalytic activity, isoelectric point, water solubility, stability, or other experimental indicators. The experimental results, after structured processing, are fed back to the self-evolving multi-agent system to adjust the dataset, reward model, reward function, or candidate sequence generation strategy. Through repeated execution of computational design and experimental validation, a closed-loop optimization is formed, ultimately obtaining engineered proteins that meet the design goals.

[0093] like Figure 2As shown, the self-evolving multi-agent system includes an orchestrator, an evaluator, a task understanding module, a model evolution module, a model asset module, an execution sandbox, and a memory system. The orchestrator receives user queries, feedback information, and task data, converting protein design goals into a typified workflow. This typified workflow includes one or more task items, each of which includes at least a task type, task instructions, structured input, expected output, task reason, resource type, and dependencies. The task type can correspond to any of the following: task construction, data processing, task adaptive pre-training, supervised fine-tuning, reward model training and evaluation, reward function assembly, reinforcement learning evolution, candidate evaluation, and task termination. The task understanding module includes a task construction sub-agent and a data processing sub-agent, used to identify task objectives, data formats, sequence fields, label fields, chain types, fixed pairing relationships, and candidate length ranges. The model evolution module includes task-adaptive pre-training sub-agents, supervised fine-tuning sub-agents, reward model training and evaluation sub-agents, reward function assembly sub-agents, reinforcement learning sub-agents, and evaluation sub-agents, which sequentially perform generator task adaptation, reward model construction, reward function generation, candidate sequence evolution, and structural and functional evaluation. The model asset module stores the base generator model, the task-adapted generator model, the evolved generator model, the reward model, and related model outputs. The execution sandbox isolates the training programs, evaluation programs, and external tools called by each sub-agent. For computationally intensive tasks requiring GPU resources, the orchestrator can limit their sequential execution based on resource availability. For task construction, data checking, and reward function assembly tasks, the orchestrator can perform early execution or parallel scheduling when there are no dependency conflicts to reduce workflow waiting time.

[0094] In this embodiment, the evaluator performs rule review and semantic review on the workflow, structured input, intermediate products, and candidate results before workflow execution and after each task is completed. Rule review may include checking whether the number of task items exceeds a preset limit, whether the input data meets a preset structure, whether the training data path and model checkpoint path are accessible, whether the task execution order conforms to dependencies, whether completed tasks are repeatedly invoked, whether the reinforcement learning task has obtained a qualified generator and reward function, and whether the model product type matches the downstream task. Semantic review may include determining whether the selected sub-agent is consistent with the user's goal, whether the data field role is consistent with the task type, whether scientific parameters require manual confirmation, whether the reward target is consistent with the design goal, and whether the current workflow should continue, be modified, or terminated. When the evaluator's review fails, the corresponding sub-agent does not enter the execution phase. The evaluator generates repair information containing error fields, error reasons, suggested repair methods, and a snapshot of the previous workflow version, and feeds this repair information back to the orchestrator. The orchestrator modifies the current workflow or regenerates the workflow based on the repair information until the workflow passes the review or reaches a preset cycle limit. After execution, each sub-agent returns a structured result containing execution status, result summary, product references, evaluation metrics, and event logs. A unified submission process writes generator products to the generator product registry, reward models that meet quality admission criteria to the reward model registry, reward functions that pass procedural testing to the reward function registry, and execution exceptions and filtering reasons to the failure history. Downstream sub-agents select only registered products from the corresponding registries, thus avoiding erroneous references due to free text descriptions and ensuring that the training, evolution, and evaluation processes can be tracked and reproduced.

[0095] like Figure 2As further illustrated, the memory system includes short-term memory, long-term memory, a memory compressor, an insight library, and filters, used to enable reflection and replanning of the self-evolving multi-agent system. Short-term memory stores user input, workflow, sub-agent call records, intermediate products, candidate sequences, reward components, evaluation metrics, and filtering reasons for the current task. When the context occupied by short-term memory reaches the preset window limit, the system truncates some historical content according to task status and information importance, and inputs the content to be retained into the memory compressor. The memory compressor extracts successful strategies, failure modes, model applicability, reward function configuration, candidate filtering reasons, and subsequent optimization suggestions from historical execution information, forming structured insights. These insights are written into the insight library and long-term memory after validity review by the filters; information that fails the review is not used as the basis for planning subsequent tasks. The orchestrator and evaluator query similar tasks, historical workflows, model products, review results, and optimization strategies in long-term memory using memory retrieval tools, and use the retrieval results to adjust the step order, sub-agent selection, reward components, candidate filtering conditions, or experimental priorities of the current task. At each candidate output stage, the system records the candidate sequence and its corresponding reward value, structural indicators, functional indicators, and exploitability indicators. When a candidate sequence is filtered out due to containing illegal amino acids, not meeting length requirements, structural confidence below the threshold, excessive repetition rate, or potential exploitability risks, the corresponding filtering reasons are written into the failure history. Thus, the orchestrator generates a workflow, the evaluator reviews it, the sub-agents execute tasks, the memory system saves the results and extracts insights, and the orchestrator re-plans based on historical information, thereby forming a sustainable reflection-evolutionary loop.

[0096] like Figure 3As shown, the reinforcement learning sub-agent employs a multi-objective group reward decoupling normalization algorithm to perform task-specific evolution of the protein sequence generator. For an input task q, the policy model generates a candidate group containing multiple candidate outputs at once. The reward function assembly sub-agent configures multiple reward components according to the task objective. These reward components may include at least one of the following: functional prediction score, binding interface confidence, global structure confidence, residue-by-residue structure confidence, sequence naturalness, sequence novelty, fixed pairing reward, amino acid validity, length constraint, and exploitability evaluation. For the same reward objective, the system first independently normalizes the reward value of each candidate output within the candidate group to obtain the objective advantage of each candidate output under that reward objective, thereby reducing the impact of differences in the numerical range and variance of different reward components on the training process. Subsequently, the system weights and combines multiple objective advantages according to preset weights corresponding to each reward objective to obtain the total advantage of each candidate output, and then performs a second normalization on the total advantage within the candidate group to obtain the final advantage used to update the policy model. During the policy model update process, the KL divergence of the policy model relative to the reference model is also calculated, and the KL divergence is used to limit the degree to which the policy model deviates from the reference model, so as to maintain the basic naturalness and training stability of the generated sequence. Unlike GDPO (Liu et al., arXiv 2026, arXiv:2601.05242.) which only uses group advantage optimization, this embodiment further introduces a repulsive KL divergence regularization term RepKL into the policy objective, and constructs a repulsive effect based on the JS divergence between candidate outputs, so that the candidate outputs maintain an appropriate distance in the representation space or probability distribution space, thereby suppressing the excessive concentration of the generated results in a few high reward regions.

[0097] To identify the technical problems of existing reinforcement learning methods in protein sequence optimization tasks, this embodiment compares ProteinRL (Sternke et al., GenBio@NeurIPS 2023.), DyNA-PPO (Angermueller et al., ICLR2020.), PPO (Schulman et al., arXiv 2017, arXiv:1707.06347.), GRPO (Guo et al., Nature 2025, 645, 633-638.), DAPO (Yu et al., Adv. Neural Inf. Process. Syst.2025, 38, 113222-113244.), GDPO, and GSPO (Zheng et al., arXiv 2025, arXiv:2507.18071.) using a unified dataset, basic generator, reward model, number of candidates, and training iteration conditions. The results are as follows: Figure 4 As shown. Figure 4 A records the reward values ​​of each algorithm during the training iteration process, which is used to evaluate the optimization process of different algorithms on the reward model objective; Figure 4 B records the policy entropy during the corresponding training process, which is used to characterize the convergence state of the generated policy; Figure 4 C compares the reward model prediction scores of the training set and the sequences generated by each algorithm to evaluate the degree of optimization of the reward objective; Figure 4 D compares the hydrophobicity distribution of the training set and the generated sequence to check whether reinforcement learning optimization causes the candidate sequence to deviate from the original data distribution; Figure 4 E calculates the repetition rate of sequences generated by each algorithm and uses the proportion of repetitive sequences to the total number of generated sequences to characterize the degree of repetition. Figure 4 F calculates the edit distance between the sequences generated by each algorithm to characterize the internal diversity of candidate sequences. The evaluation results show that ProteinRL failed to achieve stable convergence of the reward model score, and DyNA-PPO exhibited periodic oscillations during reward optimization. Although both can preserve the differences between candidate sequences to some extent, their reward optimization capabilities are limited. PPO, GRPO, DAPO, GSPO, and GDPO can improve the reward model prediction score, but the generated sequences have a high repetition rate and small edit distances, indicating that their candidate sequences are concentrated in a limited high-reward region. Therefore, existing reinforcement learning methods struggle to simultaneously address reward objective optimization and candidate sequence space exploration. Thus, a regularization mechanism that can directly adjust the distribution relationship of candidate sequences needs to be added to GDPO.

[0098] To verify the impact of different optimization strategies on the GDPO training process, this embodiment constructs the original GDPO, GDPO+D with added diversity rewards, DMA with diversity modulation advantages, DynClip with novelty-aware dynamic pruning, and RepKL with added repulsive KL divergence regularization terms. Reinforcement learning training is performed under the same base model, training data, reward model, sampling quantity, and number of iterations. The evaluation results are as follows: Figure 5 As shown. Figure 5 A records the reward model scores of each algorithm during the reinforcement learning training process. Figure 5 B records the average word-level policy entropy of each algorithm in each iteration for optimizing mini-batches. Figure 5 C statistically analyzes the reward model score distribution of the sequences generated by each algorithm after training. Figure 5 D represents the reward model scores of at most the top ten high-scoring sequences among the non-repeating sequences generated by each algorithm. Figure 5 E performs one-hot encoding on the generated sequence and uses UMAP to reduce the dimensionality, then displays the distribution of the non-repeating sequence in the low-dimensional space. Figure 5 F compares the GRAVY hydrophobicity distribution of the training set and the sequences generated by each algorithm. Figure 5 G calculates the pairwise edit distances between the unique sequences generated by each algorithm, normalized to the sequence length. Figure 5 H-statistics were used to evaluate the sequence repetition rate of each algorithm in multiple independent generation experiments. By jointly evaluating reward optimization, training entropy, candidate distribution, internal edit distance, and repetition rate using the above metrics, RepKL improved the internal edit distance of generated sequences and reduced the sequence repetition rate while maintaining the reward model score. The results indicate that adding the exclusionary KL divergence regularization term to the GDPO policy update objective enables the policy model to continue sampling different candidate sequence regions during task reward optimization, thereby alleviating the problem of candidate sequences concentrating on a few high-reward patterns.

[0099] To further characterize the impact of RepKL on the sequence space exploration process, this embodiment performs minimum spanning tree analysis on the non-repeating sequences generated by GDPO+RepKL and the original GDPO throughout the training process. The results are as follows: Figure 6 As shown. When constructing the minimum spanning tree, each node corresponds to a non-repeating candidate sequence, the node color corresponds to the number of training iterations when the sequence was generated, the edges between nodes represent the minimum connection relationship determined according to the preset sequence distance metric, and the candidate sequence generated in the last iteration is marked with red nodes. Figure 6 A and Figure 6 B shows the minimum spanning tree topology of the sequences generated by GDPO+RepKL and GDPO, respectively, which is used to present the position, connection relationship and expansion path of the candidate sequences generated at different training stages in the sequence space; Figure 6C shows the total weight of the minimum spanning tree corresponding to the two algorithms. The total weight is the sum of the distances of each connecting edge in the tree, which is used to characterize the range that the generated sequence cumulatively traverses in the sequence space. Figure 6 D illustrates the diameter of the minimum spanning tree corresponding to the two algorithms. The diameter is the maximum path length between any two nodes in the tree, representing the vertical span of the sequence exploration path. By combining training iteration information, distance relationships between candidate sequences, and the minimum spanning tree topology, the sequence discovery process of the reinforcement learning sub-agent at different training stages can be visualized, and its sequence space exploration characteristics can be quantified from both total weight and tree diameter perspectives. Figures 1 to 6 This embodiment sequentially implements the setup of computing and experimental environments, generation of typified workflows, task review and replanning, memory compression and reuse of historical experience, multi-objective reward optimization, policy update under RepKL constraints, and candidate sequence space exploration and evaluation, thus constituting a complete general self-evolving multi-agent protein design process.

[0100] Example 2: Data-Driven Protein Redesign like Figure 7 As shown in Figure A, this embodiment provides a method for protein redesign using deep mutation scanning data or sequence-activity data. This method is used to learn high-dimensional sequence-function relationships under conditions where a large number of protein sequences and their functional tags already exist, and to generate candidate protein sequences located outside the training data distribution. The system receives the wild-type sequence of the target protein and the corresponding deep mutation scanning data. The deep mutation scanning data can originate from single-point mutation, multi-point mutation, saturation mutation, or combined mutation experiments, and includes at least sample numbers, protein sequences, and functional tags. The data processing sub-agent performs format checks, missing value handling, abnormal sequence filtering, repetitive sequence handling, amino acid validity checks, sequence length statistics, and functional tag standardization on the input data. It identifies sequence columns, tag columns, sample number columns, and other auxiliary information columns, and outputs the parsed data semantic information. For input data with multiple candidate sequence columns or tag columns, the system generates structured confirmation information and requests manual confirmation, rather than inferring field functions from free text. After field confirmation, the data processing sub-agent constructs training, validation, and test sets according to a preset ratio or a hierarchical approach based on mutation sites and sequence similarity. Simultaneously, it retains the original sample numbers and the correspondence between each sample and its experimental source, enabling the model training results, candidate sequences, and subsequent experimental validation results to be traced back to the original data. Through this processing, the original deep mutation scanning data is transformed into a structured data product that can be accessed by the task-adaptive pre-training sub-agent, the reward model training and evaluation sub-agent, and the reinforcement learning sub-agent.

[0101] like Figure 7As shown in Figure B, the data-driven protein redesign process in this embodiment includes data processing, task-adaptive pre-training, reward model training, reward function assembly, reinforcement learning, and evaluation. The task-adaptive pre-training sub-agent, based on the deep mutation scan sequence of the target protein or related protein family sequences, further trains the basic sequence generator, enabling it to learn the local sequence patterns, amino acid co-occurrence relationships, mutation tolerance patterns, and sequence length distribution of the target protein. Before training begins, the system generates a training configuration including the training data path, model configuration, initial checkpoints, sequence format, training rounds, batch size, and computational resource requirements. The evaluator checks the data path, model type, sequence length, and resource conditions. If manual review is required, the training configuration is confirmed before entering the execution phase. After task-adaptive pre-training is completed, the system outputs the model configuration, vocabulary, initial checkpoints, training checkpoints, training logs, and sequence length statistics, and registers these products in the generator product registry. The reward model training and evaluation sub-agent trains a regression reward model based on the protein sequence and its functional labels to approximate the sequence-functional landscape of the target protein, and evaluates the predictive ability of the reward model using validation and test sets. The reward model must pass the quality admission criteria set by the evaluator before it can be written into the reward model registry. Reward models that fail to meet the quality admission criteria are not used for subsequent reward function assembly. The system records their training configuration, evaluation metrics, and reasons for failure, and the orchestrator determines whether to adjust the data partitioning, training parameters, or model structure before retraining. In this embodiment, the regression surrogate models trained for the four target proteins all passed the evaluator's quality admission criteria.

[0102] To illustrate the implementation of the above method in protein redesign tasks with different structural types and functional targets, this embodiment selects four target proteins: AAV2 (UniProtKB: P03135), HIS7 (UniProtKB: P06633), Kir2.1 (UniProtKB: P35561), and EcCM (UniProtKB: P0A9J8), and trains task-specific generators and regression reward models using corresponding deep mutation scan data. The functional tags for AAV2 and HIS7 are production fitness, for Kir2.1 it is membrane surface expression level, and for EcCM it is catalytic activity. The system generates redesigned sequences under uniform candidate sampling conditions, using ESM-IF1 (Hsu et al., Proc. Mach. Learn. Res. 2022, 162, 8946-8970.), ProteinMPNN (Hsu et al., Proc. Mach. Learn. Res.2022, 162, 8946-8970.), and AFDesign (Mirdita et al., Nat. Methods 2022, 19, 679-682.) as structure-information-driven control methods. The evaluator uses ESM2 log-likelihood to assess the sequence naturalness of the generated sequences; a higher log-likelihood indicates that the candidate sequences better conform to the natural sequence patterns learned by the protein sequence model. Figure 7 As shown in Figure C, the log-likelihood of the sequences generated by SEMAS and each control method for the four target proteins AAV2, HIS7, Kir2.1, and EcCM was statistically analyzed (Lin et al., Science 2023, 379, 1123-1130.). In the aggregated results for the four target proteins, SEMAS improved sequence naturalness by 29.5%, 33.9%, and 29.1% compared to ProteinMPNN, ESM-IF1, and AFDesign, respectively. No statistically significant difference was detected between SEMAS and AFDesign in the EcCM task. This evaluation was used to examine whether candidate sequences still maintain statistical characteristics consistent with natural protein sequences when reinforcement learning optimizes subsequence functional objectives, and served as a candidate selection criterion in addition to functional prediction.

[0103] like Figure 7As shown in Figure D, the evaluation agent uses a regression surrogate model with quality criteria to calculate the functional prediction score of each candidate sequence, and compares the SEMAS-generated sequence with the original training set and the top 32 sequences in the training set in terms of functional score. For AAV2 and HIS7, the functional prediction score corresponds to production fitness; for Kir2.1, the functional prediction score corresponds to membrane surface expression level; and for EcCM, the functional prediction score corresponds to catalytic activity. In the evaluation of the four target proteins, the functional prediction scores of the SEMAS-generated sequences are all higher than those of the top 32 sequences in the training set, with a 295.7% improvement in the functional prediction score in the Kir2.1 task; since Kir2.1 uses a logarithmic scale for expression indicators, this change corresponds to approximately a tenfold change in the original expression score. By comparing the SEMAS candidate sequences with the high-scoring sequences in the training set, it is possible to evaluate whether the generator merely reproduces known high-scoring patterns from the training data, or whether it can search for new candidate sequences in the sequence-functional landscape represented by the reward model. For candidate sequences whose predicted functional scores improve but whose sequence naturalness, structural confidence, or solubility-related metrics decrease, the evaluator records this as a reward bias and writes the target protein, candidate sequence, reward component, and evaluation metrics into the memory system. In the next round of the reflection-evolution cycle, the orchestrator adjusts the weights of naturalness rewards, structural proxy rewards, solubility-related rewards, or diversity rewards based on this reward bias to reduce candidate quality shifts caused by over-optimization of a single functional objective.

[0104] To evaluate the structure preservation of candidate sequences, the evaluation agent used ESMFold (Lin et al., Science 2023, 379, 1123-1130.) to predict the three-dimensional structure of candidate proteins and calculated the prediction template modeling score pTM and the prediction local distance difference test score pLDDT. pTM was used to characterize the global folding confidence of the candidate protein, and pLDDT was used to characterize the residue-by-residue structure confidence. Figure 7 E and Figure 7 As shown in Figure F, the system statistically analyzed the pTM and pLDDT of the sequences generated by SEMAS, ProteinMPNN, ESM-IF1, and AFDesign for the four target proteins. SEMAS achieved higher structure confidence in the Kir2.1 task than the compared control methods, and its structure confidence was similar to some control methods in the AAV2 and HIS7 tasks. However, in the EcCM task, its global structure confidence and residue-by-residue structure confidence were approximately 11.8% and 6.8% lower than ProteinMPNN, respectively. The system also selected representative high-scoring sequences from the candidate sets of each target protein and superimposed their predicted structures with the corresponding wild-type reference structures. Figure 7As shown in G, representative candidate sequences of AAV2, HIS7, Kir2.1, and EcCM have edit distances of 15 to 51 amino acids relative to wild-type sequences, with root mean square deviations (RMSDs) of structural superposition between their predicted structures and wild-type reference structures ranging from 0.156 Å to 1.865 Å. The evaluation agent correlates and stores the edit distance, pTM, pLDDT, and RMSD of candidate sequences with their corresponding functional prediction scores to form a candidate evaluation record that simultaneously reflects the degree of sequence variation, functional objectives, and structural preservation.

[0105] The evaluation sub-agent further assesses the secondary structure composition and predicted water solubility of sequences generated by different methods. For example... Figure 8 As shown in Figure A, the system statistically analyzes the proportions of helical, folded, and coiled structures in wild-type sequences of AAV2, HIS7, Kir2.1, and EcCM, as well as sequences generated by SEMAS, ProteinMPNN, ESM-IF1, and AFDesign, based on predicted structures. The average deviation of the secondary structure components of the SEMAS-generated sequences relative to wild-type sequences does not exceed 3.1%, and the maximum deviation of a single secondary structure component does not exceed 5.2%. In this embodiment, the secondary structure composition is used to evaluate whether the candidate sequence still maintains a local structural type distribution similar to the target protein after multiple amino acid substitutions, but it is not used as the sole basis for candidate sequence approval or rejection. The system further calculates the hydrophobicity of candidate sequences using the average hydrophobicity index GRAVY, and characterizes predicted water solubility using the negative average hydrophobicity index -GRAVY, where a higher -GRAVY value indicates higher predicted water solubility of the candidate sequence. Figure 8 B to Figure 8 As shown in Figure E, the system establishes the correlations between predicted water solubility and production fitness for AAV2 and HIS7, predicted water solubility and membrane surface expression level for Kir2.1, and predicted water solubility and catalytic activity for EcCM, respectively. The group mean and standard deviation of SEMAS and each structure-information-driven method are compared. The evaluation results, along with functional prediction scores, sequence naturalness, and structural confidence, are written into a candidate scoring table for candidate ranking, manual review, and subsequent experimental verification.

[0106] Example 3: De novo binding protein design like Figure 9As shown in Figure A, this embodiment provides a method for de novo design of binding proteins in the absence of known target-binding protein sequence functional data. The system receives target protein specifications, which include at least the amino acid sequence of the target protein and may also include a target binding region, a candidate binding protein length range, a forbidden binding region, interface residue constraints, fixed pairing sequences, and expected structural evaluation metrics. For zero-sample design tasks that only provide the target protein sequence, the orchestrator calls an existing basic sequence generator to perform cold-start candidate generation, and the reward function assembly agent automatically constructs structural proxy rewards and fixed pairing rewards based on the target sequence. For tasks that simultaneously provide known target-binding protein pairing data, the data processing agent identifies the receptor sequence, binding protein sequence, and tag sequence, and the supervisory fine-tuning agent uses the target sequence or target cues as input and the binding protein sequence as output, enabling the sequence generator to acquire target conditional generation capabilities. The candidate binding protein length can be a fixed length or within a preset length range; when the candidate length is a fixed value, the reward function uses precise length gating; when the candidate length is within an allowed range, the reward function uses length window gating. Length gating and amino acid validity gating are performed before structural proxy evaluation. Candidate sequences that do not meet the corresponding conditions are assigned low or zero scores and are not included in subsequent complex structure prediction, thereby reducing the computational resource consumption of invalid candidates.

[0107] In this embodiment, the cold start phase first generates initial candidate binding proteins based on the target protein sequence, and then uses structural proxy evaluation to determine the interface quality and structural quality of the candidate binding proteins after forming a complex with the fixed target protein. The composite reward constructed by the reward function assembler agent can include amino acid validity gating, length gating, fixed pairing reward, complex structure proxy reward, binding interface reward, binding protein monomer structure reward, sequence naturalness reward, and diversity reward. The fixed pairing reward uses the given target protein sequence as the pairing object to evaluate the matching degree between different candidate binding proteins and the same target; the complex structure proxy reward is used to characterize the interchain confidence and interface confidence after the candidate binding protein forms a complex with the target; the binding protein monomer structure reward is used to evaluate the global folding confidence and residue-by-residue structure confidence after the candidate binding protein detaches from the target; the sequence naturalness reward is used to limit the degree to which the candidate sequence deviates from the natural protein sequence pattern; and the diversity reward and RepKL regularization term are used to reduce the risk of candidate sequences being concentrated in a few similar regions. Structural proxy rewards can be used as rapid feedback during reinforcement learning training. After evolution, candidate sequences are then evaluated using AlphaFold-Multimer or other complex structure prediction methods for higher accuracy, ensuring that the computational efficiency of the reinforcement learning phase is aligned with the structural confidence required for the final candidate ranking. If the evaluator detects an increase in the interface confidence of a candidate binding protein while a decrease in monomeric pTM or pLDDT, the evaluator records this as a reward bias and returns repair information to the orchestrator, including interface indices, monomeric structure indices, and relevant candidate sequences. The orchestrator then re-invokes the reward function to assemble the sub-agent, increasing the weight of monomeric structure rewards or sequence naturalness rewards, and initiates the next round of reinforcement learning evolution. If the structural quality of a candidate binding protein meets the requirements but the interface confidence is insufficient, the system correspondingly increases the weight of fixed pair rewards or interface proxy rewards. This multi-round evaluation, feedback, and replanning balances the optimization of the binding interface with the structural stability of the binding protein itself.

[0108] like Figure 9 A and Figure 9As shown in B, this embodiment can also receive human feedback after completing the first round of cold start design and execute the second round of self-evolution based on the human feedback. The human feedback can point out the deficiencies in the interface confidence, binding site, monomer structure quality, sequence novelty, or other evaluation indicators of the candidate sequences in the previous round. The orchestrator first reads the workflow, reward function, generator version, candidate sequences, and evaluation results of the first round of tasks through a memory retrieval tool. The evaluator identifies the optimization objectives that need to be adjusted based on the human feedback, and then sequentially calls the data processing sub-agent, the reward function assembly sub-agent, the reinforcement learning sub-agent, and the evaluation sub-agent to form an updated workflow. To illustrate this process, this embodiment selects PD-L1 (UniProtKB: Q9NZQ7), CD47 (UniProtKB: Q08722), and IL-7Rα (UniProtKB: P16871) as target proteins, where PD-L1 is a monomeric receptor target, CD47 contains a flexible loop region, and IL-7Rα is a dynamic cytokine receptor. The system generates binding protein sequences under uniform candidate sampling conditions. An SFT model, obtained through supervised fine-tuning of the basic generation model based on the HuRI dataset, serves as a sequence generation control. RFdiffusion (Watson et al., Nature 2023, 620, 1089-1100.) and AFDesign (Mirdita et al., Nat. Methods 2022, 19, 679-682.) are used as structure-information-driven control methods. The evaluation sub-agent uses AlphaFold-Multimer to predict the complex structure formed by the candidate binding protein and the target protein, and calculates the inter-chain prediction template modeling score (ipTM), interface-specific confidence score (ipSAE) based on binding interface residues, residue-by-residue structure confidence score (pLDDT), and global folding confidence score (pTM). The edit distance between the candidate binding protein and its nearest natural homologous sequence is used as an indicator of sequence novelty. Figure 9 B to Figure 9As shown in Figure D, the system compares the evaluation results of ipTM, ipSAE, and sequence novelty of binding proteins generated by SEMAS, SFT, RFdiffusion, and AFDesign for CD47, IL-7Rα, and PD-L1. Compared with the SFT model, SEMAS improved the ipTM, ipSAE, pLDDT, pTM, and sequence novelty of the generated sequences by 160%, 243%, 40%, 54%, and 108%, respectively; compared with RFdiffusion, SEMAS improved the ipTM and ipSAE of the generated sequences by 51% and 62%, respectively, and the sequence novelty by 16%; compared with AFDesign, SEMAS improved the ipTM and ipSAE of the generated sequences by 160% and 159%, respectively, and the sequence novelty by 42%. Figure 9 E and Figure 9 F further illustrates the pTM and pLDDT of the binding protein generated by different methods, with SEMAS showing the highest global folding confidence and residue-by-residue structure confidence on two of the three targets, and in the same range of metrics as the structure information-driven methods on the remaining targets.

[0109] like Figure 9 As shown in G, this embodiment selects representative binding proteins from candidate sequences generated by SEMAS for CD47 and displays the predicted complex structures formed by these proteins with CD47. The CD47 target protein of the complex is indicated in pink, and the candidate binding proteins are indicated in blue. A magnified area shows the interfacial contact residues between the candidate binding protein and CD47, and the corresponding ipTM and ipSAE of the complex are labeled. The system can evaluate the complex structures of IL-7Rα and PD-L1 candidate binding proteins in the same manner. Figure 10 B and Figure 10 As shown in C, the representative IL-7Rα binding protein and PD-L1 binding protein generated in the second round of self-evolution form complexes with their respective targets. The binding interface and candidate binding protein residues involved in the interaction are marked in the magnified local image, and the ipTM and ipSAE of the corresponding complexes are recorded. Figure 10 A uses a radar chart to illustrate the normalized results of the binding proteins designed in the first and second rounds of SEMAS on five metrics: ipTM, pLDDT, pTM, ipSAE, and sequence novelty. The second round showed a 50.9% improvement in ipTM, a 2.0% improvement in pLDDT, a 7.9% improvement in pTM, a 15.1% improvement in ipSAE, and a 17.2% improvement in sequence novelty compared to the first round. The candidate sequences, reward function configurations, model versions, and evaluation metrics from both the first and second rounds were written into a memory system to record the optimization directions triggered by manual feedback and their corresponding results.

[0110] To evaluate the relationship between structural quality and sequence novelty of candidate binding proteins across the entire generated sequence space, rather than just a few representative complexes, this embodiment performs population-level correlation analysis on all candidate binding proteins generated by SEMAS, RFdiffusion, and AFDesign. Figure 10 D to Figure 10 As shown in Figure F, the system establishes scatter plots between the sequence novelty of sequences generated by three methods and the residue-by-residue structural confidence (pLDDT) of binding proteins, and calculates Pearson correlation coefficients through linear regression fitting. The correlation coefficients (r) for SEMAS, RFdiffusion, and AFDesign are -0.079, -0.272, and -0.894, respectively. Figure 10 G to Figure 10 As shown in Figure I, the system further establishes a scatter plot relationship between sequence novelty and the global folding confidence (pTM) of the binding protein, where the Pearson correlation coefficients (r) for SEMAS, RFdiffusion, and AFDesign are 0.561, -0.272, and -0.475, respectively. This correlation analysis is used to characterize the changing trends of monomeric residue-by-residue structural confidence and global folding confidence when candidate sequences deviate from their natural homologous sequences. The evaluation agent associates and stores the novelty, pLDDT, pTM, ipTM, and ipSAE of each candidate sequence, and writes the population correlation as the candidate set evaluation result into the memory system so that the evaluator can determine whether the current reward function causes the structural confidence to continuously decrease as sequence novelty increases.

[0111] Example 4: De novo antibody design This embodiment provides a de novo antibody design method based on a self-evolving multi-agent system. This method generates candidate sequences for the variable regions of the heavy and light chains of antibodies based on the target antigen sequence, even when no known antigen-antibody pairing sequence or reference antibody sequence is available. Antibody molecules have larger molecular sizes and more complex multi-chain structures compared to single-chain binding proteins, and their complementarity-determining regions (CDRs) exhibit high conformational flexibility. Therefore, candidate antibodies not only need to form a binding interface that matches the target antigen but also need to maintain the folding stability of both the heavy and light chains and the quality of the VH-VL interchain pairing. The antigen specifications received by the system include at least the amino acid sequence of the target antigen and may also include the expected binding region, the forbidden binding region, the antibody chain format, the candidate length window, the fixed frame sequence, CDR constraints, and exploitability constraints. The data processing sub-agent or task construction sub-agent identifies the antigen information, heavy chain sequence format, light chain sequence format, interchain separators, CDR location, and candidate length range, and generates confirmed data semantic information. Antibody chain formats can be represented as paired heavy and light chains, or as single-chain antibodies linked by linking peptides. For paired representations, the system retains information on the heavy and light chains and interchain separators; for single-chain antibody formats, the system retains information on linking peptides or intrachain separators. Before training or evolution begins, the evaluator checks whether the antigen sequence, antibody chain format, and length window have been confirmed. If necessary information is missing, it prevents reward function assembly and reinforcement learning execution.

[0112] In this embodiment, the system first adapts the basic sequence generator to the antibody sequence data to form an initial antibody generator suitable for subsequent antigen conditional evolution. Specifically, the task-adaptive pre-training sub-agent or the supervised fine-tuning sub-agent trains the basic sequence generator using the Observed Antibody Space dataset, enabling it to learn the amino acid distribution, framework region regularity, complementarity-determining region composition, chain length distribution, and VH-VL pairing relationship of the natural antibody heavy and light chains. The training data is represented by a unified pairing sequence consisting of heavy chain sequences, light chain sequences, and inter-chain separators, while retaining the boundary information of each chain, allowing the model to distinguish between heavy and light chains during generation. After training, the system performs training set similarity checks and nearest natural homologous sequence searches on the generated sequences to check whether the generator only reproduces known sequences from the training data. The trained generator is registered in the generator product registry and serves as the initial policy model for reinforcement learning evolution. Figure 11 A and Figure 11As shown in Figure B, upon receiving the target antigen sequence, the orchestrator generates a de novo antibody design workflow that includes data processing, reward function assembly, reinforcement learning, and evaluation. This allows the generator, which has already learned the patterns of antibody sequences, to produce paired VH-VL candidate sequences based on antigen cues. This workflow does not require input of known positive antibody sequences; antigen-related selection pressure is provided by the reward function and complex structure evaluation, thus enabling zero-sample antibody design targeting the target antigen. The reward function assembly sub-agent constructs an antibody structure funnel and a multi-objective reward function for the antibody multi-chain design task. The antibody structure funnel processes candidate antibodies in the following order: candidate sequence extraction, heavy and light chain analysis, amino acid validity checking, chain length checking, frame region and complementarity-determining region checking, sequence naturalness evaluation, antibody self-structure evaluation, antigen-antibody complex evaluation, and candidate diversity evaluation. For candidate sequences that do not meet the requirements for amino acid validity, chain format, or length, the system assigns a low or zero score through pre-gating to prevent them from entering the computationally expensive structure prediction stage. The multi-objective reward function may include at least one of the following: antigen-antibody chain confidence, binding interface confidence, antibody global structure confidence, residue-by-residue structure confidence, VH-VL pairing confidence, VH-VL interface confidence, sequence naturalness, and sequence novelty. It may also incorporate surface hydrophobicity, surface positive charge, surface negative charge, charge symmetry, complementarity-determining region length, and other therapeutic antibody profiling indicators.

[0113] To illustrate the above-mentioned de novo antibody design method, this embodiment selects seven antigens as target antigens: H1N1 (PDB: 3GBN), Spike HR2, VEGF-A (PDB: 2FJG), HEWL (PDB: 1MLC), HER2 (PDB: 1N8Z), Integrin-α-1 (PDB: 1MHP), and Ang2 (PDB: 4ZFG). These antigens cover small globular enzymes, receptor proteins, viral fusion-related proteins, and other different topological structures. The system generates paired VH-VL candidate sequences for each antigen under uniform candidate sampling conditions, with MAGE (Wasdin et al., Cell 2025, 188, 7206-7221.e16.), DiffAb (Luo et al., Adv. Neural Inf. Process. Syst. 2022, 35, 9754-9767.) and RFAntibody (Watson et al., Nature 2023, 620, 1089-1100.) as control methods. MAGE represents the sequence generation method, DiffAb represents the sequence-structure co-design method, and RFAntibody represents the method of generating the backbone first and then sampling the sequence. IgFold (Ruffolo et al., Nat. Commun. 2023, 14, 2389), IgLM (Shuai et al., Cell Syst. 2023, 14, 979-989.e4), AntiBERTy (Ruffolo et al., arXiv 2021, arXiv:2112.07782), and AbLan (Olsen et al., Bioinform. Adv. 2022, 2, vbac046) were used for antibody structure prediction or antibody sequence characterization, but they do not support de novo generation of antigens based on given antigen conditions, and therefore were not included in the unified generation comparison in this embodiment. For each method and each antigen, the antigen-VH-VL complex structure was predicted using AlphaFold-Multimer. Complex evaluation included ipTM for characterizing overall confidence between antigen and antibody chains, ipSAE for characterizing the quality of the antigen-antibody binding interface, pLDDT for characterizing residue-by-residue structural confidence, and pTM for characterizing global folding confidence. Sequence novelty was calculated using edit distances between candidate antibody VH and VL and their nearest natural homologs, and the results for both heavy and light chains were summarized.

[0114] like Figure 11 C to Figure 11As shown in F, the evaluation sub-agents compared the ipTM, ipSAE, pLDDT, and pTM of SEMAS, MAGE, DiffAb, and RFAntibody against seven antigens. For the overall interchain confidence ipTM, SEMAS achieved high values ​​on almost all antigen targets, with its average value being 22.1%, 14.8%, and 9.4% higher than MAGE, DiffAb, and RFAntibody, respectively. For the binding interface confidence ipSAE, SEMAS achieved the highest value on four antigen targets, and the difference from the corresponding highest value on the other three antigen targets did not exceed 0.03. For the residue-by-residue structure confidence pLDDT, SEMAS achieved the highest average value on all seven antigen targets, with improvements of 8.5%, 4.1%, and 4.0% compared to MAGE, DiffAb, and RFAntibody, respectively. For the global folding confidence level (pTM), SEMAS achieved the highest average value on six of the seven antigen targets, with a pTM of 0.901 for Spike HR2 and 0.909 for DiffAb. The evaluation results were based on a statistical design of 32 antibodies generated for each target using each method, and the differences between SEMAS and the control methods were compared using a two-sided Mann-Whitney U test corrected for Benjamini-Hochberg false discovery rate. The system entered the antigen sequence, VH sequence, VL sequence, ipTM, ipSAE, pLDDT, pTM, and statistical grouping information for each candidate antibody into the candidate evaluation table.

[0115] To separate the evaluation of antigen-antibody complexes from the evaluation of antibody self-folding, this embodiment also extracts the antibody heavy and light chains from the antigen-VH-VL complex and uses AlphaFold-Multimer to predict the antibody's self-structure without the antigen. Antibody self-folding evaluation includes ipTM for characterizing VH-VL pairing quality, ipSAE for characterizing VH-VL interface quality, pTM for characterizing the overall antibody structural quality, pLDDT for characterizing antibody residue-by-residue structural confidence, and ipLDDT for characterizing VH-VL pairing interface residue-by-residue confidence. Figure 12 A to Figure 12As shown in Figure E, the system uses a radar chart to record the five indicators of antibody generation for seven antigens by SEMAS, MAGE, DiffAb, and RFAntibody. Each axis represents an antigen, and the radial distance represents the average indicator of the corresponding method on that antigen. By predicting the antigen-antibody complex structure and the antibody's own structure separately, candidate sequences that only have a high interface score in the complex evaluation but exhibit VH-VL pairing abnormalities or decreased monomer folding confidence after detachment from the antigen can be identified. In the evaluation results of this embodiment, SEMAS is in a high range for antibody's own structure indicators for most targets and does not show VH-VL domain folding failure. In the control methods, there are cases where the local or global structure confidence is significantly reduced. For example, the pLDDT of MAGE antibody generation against VEGF-A is reduced to 69.7, and the pLDDT and pTM of DiffAb antibody generation against HEWL are 78.1 and 0.550, respectively.

[0116] In this embodiment, the evaluation sub-agent also employs Pareto front analysis to jointly evaluate antibody structure confidence and sequence novelty. For example... Figure 13 As shown in Figure A, the system calculates the average ipTM and average sequence novelty of antibodies generated against each antigen by SEMAS, MAGE, DiffAb, and RFAntibody, and represents each combination of method and antigen as a data point in the ipTM-sequence novelty space. Figure 13 As shown in Figure B, the system calculates the Pareto front hypervolume of each method in the two-dimensional space based on a unified reference point, where the hypervolumes of SEMAS, RFAntibody, DiffAb, and MAGE are 0.220, 0.192, 0.164, and 0.147, respectively. Figure 13 As shown in C, the system counts the number of times each method enters the Pareto front of ipTM-sequence novelty. Among them, SEMAS accounts for 55% of all Pareto front occurrences, DiffAb and MAGE each account for 18%, and RFAntibody accounts for 9%. Figure 13 D to Figure 13 F uses the same method to establish the ipSAE-sequence novelty space, and calculates the data point distribution, Pareto hypervolume, and the number and proportion of entries into the Pareto front for each method in this space. Pareto analysis does not require pre-combining structural confidence and sequence novelty into a single weighted score, but rather retains candidate sequences that have no other candidates that are comprehensively superior to it for at least one objective, thus forming a candidate set that takes into account both the quality of the antigen-binding interface and the degree of sequence exploration.

[0117] Example 5: Wet experimental data-guided antibody design like Figure 14As shown in Figure A, this embodiment provides a method for guiding iterative antibody design using wet experimental data or deep mutation scanning data. The first round of design is a cold-start design. When the system only receives the target antigen sequence and no corresponding antibody experimental data is available, it generates paired heavy chain variable region and light chain variable region candidate sequences according to the de novo antibody design process described in Example 4. After the first round of candidate antibodies undergoes evaluations of sequence validity, chain type, candidate length, antigen-antibody complex structure, antibody self-folding structure, and exploitability, a recommended experimental candidate list is formed. The recommended experimental candidate list includes at least the candidate number, antigen identifier, heavy chain sequence, light chain sequence, generation round, reward component, ipTM, ipSAE, pTM, pLDDT, VH-VL pairing index, sequence novelty, exploitability index, and recommended experimental priority. The recommended experimental priority can comprehensively consider predicted binding quality, complex structure confidence, antibody self-folding quality, candidate sequence diversity, sequence novelty relative to natural antibodies, therapeutic antibody profiling risk, and synthesis difficulty, allowing a limited number of candidates to be prioritized for experimental validation. The first round of candidate antibodies can be directly synthesized, constructed, expressed, purified, and functionally detected. They can also be used to construct deep mutation scanning libraries to obtain candidate sequences and their corresponding experimental readings. Wet experimental data, as an external feedback source for the self-evolving multi-agent system, does not directly modify the functional definitions of the orchestrator, evaluator, or individual sub-agents. Instead, it is first converted into structured feedback data, which is then input into the next round of workflow by the data processing sub-agent.

[0118] Wet assay validation may include at least one of the following: candidate antibody sequence synthesis, protein expression vector construction, cell transfection or microbial transformation, antibody expression, antibody purification, binding assay, affinity assay, stability assay, aggregation assay, cell binding assay, and functional activity assay. Binding assays may employ enzyme-linked immunosorbent assay (ELISA), biolayer interference assay (BIA), surface plasmon resonance assay (SPR), or cell binding assay. Affinity assays may obtain equilibrium dissociation constant, binding rate constant, and dissociation rate constant. Stability and exploitability assays may include thermal stability assays, size exclusion chromatography (SOCCI), aggregation ratio assays, surface hydrophobicity assays, or other therapeutic antibody profiling-related assays. Experimental results are compiled into structured feedback data, which may include at least one of the following: candidate number, heavy chain sequence, light chain sequence, antigen identifier, expression level, purification yield, binding signal, affinity parameter, binding rate, dissociation rate, thermal stability, aggregation ratio, cell binding signal, functional activity, experimental batch, reason for experimental failure, and positive label. If a candidate antibody fails to express, the system records the failure and its corresponding sequence as negative feedback. If a candidate antibody has a strong binding signal but also a high aggregation ratio, the system saves its binding experiment results and exploitability risk separately, rather than simply classifying the candidate as positive or negative. Thus, different types of experimental readings can be converted into multiple independent learning objectives for use in the next round of reward model training and reward function assembly.

[0119] The data processing sub-agent performs sequence format checks, candidate number matching, experimental batch correction, missing value handling, outlier identification, label conversion, and training data partitioning on the structured wet experimental feedback data, generating a dataset that can be used by the reward model training and evaluation sub-agent. When experimental data comes from deep mutation scans, the data processing sub-agent can convert it into continuous regression labels, discrete classification labels, or candidate ranking labels based on the properties of the experimental readings. For cases where multiple experimental indicators exist for the same antigen, the system saves expression, binding, affinity, stability, and exploitability labels separately, and records the unit, detection method, and batch information for each label. The reward model training and evaluation sub-agent selects regression, classification, or ranking training tasks based on the label type and evaluates the reward model using validation and test sets. Only reward models that pass the evaluator's quality admission can enter the reward model registry and participate in the next round of reward function assembly. When the quantity and quality of experimental data meet the training requirements, the system can train affinity reward models, expression reward models, or other experimental indicator reward models separately; when the experimental data is insufficient to support multiple independent models, a single comprehensive reward model can also be constructed based on the confirmed labels. Reward models that fail to pass quality admission must not be used for reinforcement learning. The reasons for failure, data size, validation metrics, and training configuration are written into the memory system, and the orchestrator decides whether to retrain or switch to the alternative optimization strategy described in Example 6.

[0120] like Figure 14 As shown in Figure B, in the second round of data-guided antibody design, the orchestrator sequentially calls the data processing sub-agent, the task adaptive pre-training sub-agent, the reward model training and evaluation sub-agent, the reward function assembly sub-agent, the reinforcement learning sub-agent, and the evaluation sub-agent. The task adaptive pre-training sub-agent further adapts the antibody generator using well-organized deep mutation scan sequences or wet experimental candidate sequences, enabling it to learn the local sequence distribution sampled experimentally in the current antigen task. The reward function assembly sub-agent combines the experimental data reward model with quality-admission criteria, amino acid legitimacy gating, heavy and light chain length constraints, antibody structure funnel, antigen-antibody interface rewards, VH-VL pairing rewards, sequence naturalness rewards, diversity rewards, and therapeutic antibody profiling proxy rewards to form the second-round multi-objective reward function. The reinforcement learning sub-agent performs antibody sequence evolution according to the multi-objective group reward decoupling normalization method described in Example 1, and maintains the distribution differences between candidate sequences through RepKL regularization. The first round of cold-start model, the Top32 wet experimental high-quality sequences, and the second round of SEMAS (+Data) generated sequences were stored as three different candidate sets to evaluate the sub-agents by comparing affinity predictions, exploitability indicators, and changes in sequence composition before and after the introduction of experimental data. The second round of generated sequences still needed to undergo antigen-antibody complex evaluation and antibody self-folding evaluation to prevent the experimental reward model from pulling candidate sequences to regions with high experimental prediction values ​​but insufficient structural quality.

[0121] like Figure 14 As shown in Figure C, this embodiment compares the affinity prediction values ​​of antibodies designed using training set sequences, the top 32 sequences selected through wet screening, and SEMAS(+Data) for five antigens: HEWL, VEGF-A, H1N1, Ang2, and Spike HR2. Each bar represents the average value of the corresponding candidate set, the error bar represents the standard deviation, and the scatter plot represents a single candidate sequence. The changes in the affinity prediction values ​​of SEMAS(+Data) relative to the top 32 sequences for the above five antigens are 1.2%, 3.1%, 4.2%, 10.7%, and 0.2%, respectively. Figure 14D shows the therapeutic antibody profiles of the first-round cold-start SEMAS, Top32, and SEMAS(+Data) candidate antibodies using the TAP (Raybould et al., Proc. Natl. Acad. Sci. USA 2019, 116, 4025-4030.) pass rate. TAP pass indicates that the candidate antibody simultaneously meets all selected exploitability criteria. The bars represent the average pass rate on different antigens, the error bars represent the standard deviation, and the connected data points represent the pass rate for each antigen. The average TAP pass rate for SEMAS(+Data) is 5.7% higher than that of Top32 and 19.5% higher than that of cold-start SEMAS. Figure 14 The significance of C was evaluated using a two-sided Mann-Whitney U test corrected for Bonferroni. Figure 14 The number of candidates in each candidate set in D that passed and failed the TAP test was evaluated using a Bonferroni-corrected chi-square test. The system associates and saves the affinity prediction results, TAP pass status, and various risk indicators with the candidate sequences to generate a recommended list for the next round of experiments.

[0122] like Figure 14 E to Figure 14 As shown in H, the evaluation sub-agent further compares the distribution of therapeutic antibody profiling indicators among clinical reference antibodies, first-round cold-start SEMAS candidates, Top32 sequences, and SEMAS (+Data) candidates. Figure 14 E shows the hydrophobic distribution on the antibody surface. Figure 14 F shows the distribution of negative charge on the antibody surface. Figure 14 G indicates the distribution of positive charge on the antibody surface. Figure 14 H shows the distribution of charge symmetry parameters for the antibody heavy and light chains, and the red dashed lines in each figure represent the risk thresholds of the corresponding exploitability indicators. For example... Figure 15 As shown in A, the system also compares the total complementarity determination region (CDR) length distributions of the first-round cold-start SEMAS, Top32, SEMAS(+Data), and clinical reference antibodies, and uses the risk threshold defined by the Therapeutic Antibody Profile (TAP) to indicate the allowable range of the total CDR length. Figure 15 The percentages to the right of A represent the proportion of sequences in each candidate set that fall within the threshold range, with the proportions corresponding to cold-start SEMAS, Top32, SEMAS (+Data), and clinical reference antibody being 96%, 100%, 100%, and 91%, respectively. Figure 15 B shows the overall amino acid frequencies of the cold-start SEMAS, Top32, and SEMAS(+Data) candidate antibodies across the full sequence range, used to examine whether data-guided evolution alters the overall amino acid composition of the candidate set. Figure 15 C to Figure 15 H shows the amino acid composition distribution of the three candidate sets in CDRH1, CDRH2, CDRH3, CDRL1, CDRL2, and CDRL3, respectively. The coordinate axis of each radar chart corresponds to a standard amino acid, and the radial distance represents the frequency of occurrence of the amino acid in the corresponding complementarity-determining region.

[0123] Example 6: Alternative Data-Driven Optimization Strategy This embodiment provides an alternative data-driven optimization method for use when the reward model fails to meet quality criteria or when there is insufficient experimental data. After the reward model training and evaluation sub-agent completes training, the evaluator determines whether the reward model meets the quality criteria based on preset validation metrics, test metrics, data size, and prediction stability. For continuous experimental readings, the quality criteria can be based on rank correlation coefficient, mean squared error, or other regression metrics; for classification labels, the quality criteria can be based on macro-average F1 score, area under the receiver operating characteristic curve, or other classification metrics. When the reward model fails to meet the quality criteria, the evaluator prevents the reward model from entering the reward model registry, the reward function assembly sub-agent cannot reference the unqualified model, and the reinforcement learning sub-agent cannot update the antibody generator based on its output. The evaluator generates repair information including data size, label distribution, training configuration, validation results, and reasons for failure. The orchestrator selects to re-divide the data, adjust the training configuration, request supplementary experimental data, or switch to proxy rewards based on the repair information. Alternative strategies can be automatically triggered by the evaluator or specified by the user through structured confirmation information. When the evaluator's suggestion is inconsistent with the user-specified strategy, the system requests manual confirmation before execution to avoid automatically starting high-computation training or adopting unconfirmed optimization objectives when there is insufficient data or the model is unreliable.

[0124] The proxy reward used in this embodiment is not a simulated reward, a placeholder reward, or a default value returned after model failure, but a verifiable reward component with a clear input, calculation process, and output meaning. The reward function assembler agent can select at least one of the following based on task conditions: structural proxy reward, molecular docking score proxy reward, experimental high-quality sample distribution-driven reward, sequence naturalness reward, and exploitability proxy reward. It also registers the input format, output key, calculation parameters, weights, gating rules, and applicable scope for each reward component. After assembly, the proxy reward needs to undergo candidate sequence parsing testing, illegal amino acid testing, length boundary testing, empty input testing, output key consistency checking, and numerical range checking. Only proxy rewards that pass these tests can be written into the reward function registry. When experimental data for a certain antigen cannot support a stable reward model but can provide a small number of high-quality candidates, the system can construct a distribution-driven reward function. Based on the sequence distance or characterization distance between the generated candidate and the experimental high-quality sample, the system guides the generated sequence to move towards the region where the high-quality sample is located. The distribution-driven reward can also combine the naturalness of the candidate sequence relative to the training distribution, amino acid legality, heavy and light chain lengths, and antibody structure proxy score to avoid simply replicating a small number of experimental positive samples.

[0125] like Figure 16 As shown in Figure A, in the data-guided optimization task for the HER2 antibody, the system sequentially performs data processing and task adaptive pre-training, and attempts to train a reward model using wet experimental data. When the reward model fails to pass the evaluator's quality threshold after training, the evaluator marks the task status as a reward model training failure, preventing the reward model from entering the reward function assembly process, and returns policy switching information to the orchestrator. The orchestrator, based on this information, calls the reward function assembly sub-agent, combining the proxy reward based on antibody-antigen molecule docking score with antibody structure confidence, antigen-related interface confidence, sequence naturalness, and diversity rewards to form an alternative multi-objective reward function for the HER2 task. The reinforcement learning sub-agent uses this reward function to generate the next round of candidate antibodies, and the evaluation sub-agent then uses a unified structure prediction method to evaluate the ipTM, ipSAE, pTM, and pLDDT of the candidate antibodies. Figure 16 As shown in B, the system uses radar charts to compare the results of the first round of cold-start SEMAS, Top32 wet-test screening sequences, and SEMAS (+Data) candidate antibodies on the above four indicators. Figure 16As shown in Figure C, the system displays representative antibody-antigen complexes formed by Top32 and SEMAS (+Data) against HER2, identifies residues in the VH-VL domain that interact with HER2, and records the corresponding candidate edit distances (ED), ipSAE, and ipTM. The ipSAE and ipTM of the Top32 candidate are 0.521 and 0.62, respectively, while the edit distance of the SEMAS (+Data) candidate is 6, and the ipSAE and ipTM are 0.973 and 0.87, respectively.

[0126] like Figure 16 As shown in Figure D, in the data-guided optimization task for the Integrin-α-1 antibody, due to the insufficient number of available experimental candidates to support stable reward model training, the orchestrator adopts a simplified data-guided workflow. Instead of performing task-adaptive pre-training and reward model training, it sequentially calls the data processing sub-agent, the reward function assembly sub-agent, the reinforcement learning sub-agent, and the evaluation sub-agent. The data processing sub-agent identifies high-quality candidate sequences and their complementary determinant regions from limited experimental data. The reward function assembly sub-agent uses these high-quality experimental sequences as distribution-guided targets and combines the distance between candidate sequences and target sequences, antibody sequence naturalness, structural confidence, interface confidence, and candidate diversity into a simplified reward function. Figure 16 E uses a radar chart to illustrate the first round of cold start SEMAS, Top32, and SEMAS (+Data) for Integrin-α-1, including ipTM, ipSAE, pTM, and pLDDT / 100. Figure 16 F shows representative antibody-antigen complexes formed by Top32 and SEMAS(+Data) with Integrin-α-1, and labels the residues involved in antigen interaction, edit distance (ED), ipSAE, and ipTM in the antibody. The ipSAE and ipTM of the Top32 candidate are shown in the figure as 0.003 and 0.46, respectively, while the edit distance of the SEMAS(+Data) candidate is 75, and the ipSAE and ipTM are 0.980 and 0.83, respectively. These structural metrics serve as a unified evaluation result after alternative optimization, but are not directly used as a replacement for the quality admission rules of the reward model.

[0127] During the alternative optimization process, the evaluator still performs antigen-antibody complex evaluation, antibody self-folding evaluation, and exploitability evaluation separately. Candidates with higher agent rewards are not automatically added to the recommended experimental candidate list. If a candidate antibody has a high docking score but the pairing quality of ipTM, ipSAE, or VH-VL is below the preset conditions, the evaluator writes the filtering reason into the failure history. If candidates generated by distribution traction are too close to high-quality experimental samples and lead to an increase in the candidate set repetition rate, the evaluator can request an increase in the weight of the RepKL regularization term or diversity reward. If a candidate sequence has high structural confidence but surface hydrophobicity, charge patches, or charge symmetry exceeds the risk threshold, the system can add the exploitability agent to the next round of reward function. The candidate sequences, agent reward types, reward parameters, structural indicators, and experimental results obtained from alternative optimization are all written into the memory system. As subsequent wet experimental data gradually increases and reaches the training requirements of the reward model, the orchestrator can terminate the simplified workflow, re-call the reward model training and evaluation sub-agent, and switch to the complete data guidance process described in Example 5 after the reward model passes the quality admission, so that the system can adjust the optimization path as the data scale and quality change.

[0128] Implementation details of Examples 1-6 1. Evolutionary Algorithm 1.1 GDPO-RepKL with repulsive KL divergence To address the pattern collapse problem and encourage reinforcement learning frameworks to generate diverse candidate sequences, this invention introduces the GDPO algorithm with repulsive KL divergence, namely GDPO-RepKL. The RepKL optimizer minimizes a joint objective function, which includes a pruning agent policy loss, a KL regularization term to constrain the deviation of the policy model from the reference model, and pairwise repulsive JS divergence terms to improve the diversity of sequences within candidate groups.

[0129] in, Used to control the strength of KL regularization terms. Used to control the strength of the exclusionary diversity term. Policy probability ratio. Defined as:

[0130] Advantages of using GDPO-normalized loss reduction for pruning agents Perform the calculation:

[0131] Following the GDPO calculation method, the reward components are first normalized within each sampled candidate group. Then, the normalization advantages corresponding to different reward objectives are weighted and summed, and standardized over the entire training batch. For the m-th reward component of the j-th candidate sequence, its intra-group normalization advantage is defined as:

[0132]

[0133]

[0134] in, Let G represent the reward value of the j-th candidate sequence under the m-th reward objective, and let G represent the size of the candidate group corresponding to the same input prompt. This represents a preset constant used to improve numerical stability.

[0135] To regularize the update magnitude of the policy model, this invention employs a low-variance k3 estimator to approximate the KL divergence between the policy model and the reference model. The KL divergence estimate at the $t$-th word position is:

[0136] At valid lexical positions that are not part of the filler, the sequence-level average KL divergence is:

[0137] in, `t` is an indicator function for non-filled tokens; it takes a value of 1 when the t-th position is a valid generated token, and a value of 0 otherwise. Let... and Let represent the word-level policy probability distributions of the i-th and k-th generated sequences within the candidate group at position t, respectively. For valid positions shared by the two sequences, a mixed probability distribution is first defined:

[0138] The Jensen-Shannon divergence at position t is symmetrically defined as:

[0139] Among them, each KL divergence is within the complete vocabulary space. Calculations are performed on the above:

[0140] To address the issue of inconsistent generated lengths among different candidate sequences, JS divergence is calculated only at positions where both candidate sequences contain valid tokens. The set of valid positions shared by the i-th and k-th sequences is defined as follows:

[0141] The pairwise JS divergence between the i-th sequence and the k-th sequence is defined as:

[0142] Finally, by averaging the sequence-level JS divergences of all non-repeating sequence pairs within the candidate group, the candidate group-level repulsive JS divergence term is obtained:

[0143] Where B represents the number of candidate sequences used to calculate the repulsive JS divergence. This is because the joint objective function introduces a negative sign. In minimizing the overall loss function, the optimizer tends to increase the JS divergence between different candidate sequences within the group, thereby prompting the policy model to explore mutually distinct sequence distributions and reducing the risk of generated sequences being concentrated in a few high-reward patterns.

[0144] 1.2 GDPO+D with Diverse Rewards To address the issues of the original GDPO algorithm, such as the tendency for candidate sequences generated during training to concentrate in a few high-reward regions, high repetition within the same candidate group, and insufficient sequence space exploration, a diversity reward is introduced at the candidate group level of the GDPO algorithm, forming a GDPO+D algorithm with a diversity reward. For a candidate group containing G candidate sequences generated from the same prompt, the repetition between each candidate sequence and other candidate sequences within the same candidate group is determined. When a candidate sequence is identical to at least one other candidate sequence within the candidate group or meets a preset high similarity condition, the diversity reward for that candidate sequence is set to 0; when the candidate sequence is not identical to any other candidate sequence within the candidate group and does not meet the high similarity condition, the diversity reward for that candidate sequence is set to 1. For the j-th candidate sequence generated for the i-th question... Its diversity reward is defined as:

[0145] in, Let represent the diversity reward of the j-th candidate sequence corresponding to the i-th question. Let represent the j-th candidate sequence corresponding to the i-th question, and k represent the index of other candidate sequences within the same candidate group that are different from j. This indicates that two candidate sequences are identical or meet a preset high similarity condition. This represents the indicator function. The high similarity condition can be determined based on whether the candidate sequences are completely identical, whether the sequence consistency reaches a preset threshold, whether the edit distance is lower than a preset threshold, or other sequence similarity indicators. Through the above settings, duplicate or highly similar candidate sequences within the same candidate group do not receive diversity rewards, while candidate sequences that differ from other candidate sequences receive diversity rewards. To reduce the impact of differences in diversity reward distribution within different candidate groups on strategy updates, diversity rewards within the same candidate group are normalized. The normalized diversity advantage of the j-th candidate sequence corresponding to the i-th question is calculated as follows:

[0146] in, This represents the normalized diversity advantage of the j-th candidate sequence corresponding to the i-th question. This represents the average diversity reward within the same candidate group. Let G represent the standard deviation of the diversity reward within the same candidate group, and let G represent the number of candidate sequences corresponding to the same cue. This represents a preset constant used to avoid zero denominators and improve numerical stability. By normalizing within candidate groups, the diversity advantage reflects the degree of difference between each candidate sequence and other candidate sequences in the same group, and allows this diversity advantage to participate in subsequent advantage calculations along with correctness, format, functionality, structure, or other task advantages. Before global batch normalization, the normalized diversity advantage is combined with other task advantages and expressed using coefficients. The weight of diversity exploration signals in the overall advantage is controlled. The combined advantage of the j-th candidate sequence corresponding to the i-th question is calculated as follows:

[0147] in, This represents the combined advantage of the j-th candidate sequence corresponding to the i-th question. This represents the set of task objectives involved in the advantage calculation. This indicates the advantage of the candidate sequence under the nth task objective. This represents the weighting of diversity advantages. The task objectives may include functional scores, structural confidence, combined interface quality, sequence validity, format compliance, or other preset optimization objectives. The combined advantages are then globally batch normalized according to the GDPO algorithm and used for parameter updates in the policy model.

[0148] 1.3 GDPO-DMA with Diverse Modulation Advantages In one implementation, after the GDPO algorithm completes intra-group reward normalization, multi-objective advantage combination, and global batch normalization, the obtained basic advantages are further modulated based on the diversity of candidate sequences to form a GDPO algorithm with diversity modulation advantages, i.e., the DMA algorithm. Let Indicates training batch The hint index in the middle makes This represents the sequence index in a candidate group containing G candidate sequences. The fundamental advantage obtained after global batch normalization in step 3 of the GDPO framework is represented as follows: To quantify the uniqueness of each generated sequence, a diversity score is assigned to the j-th candidate sequence corresponding to the i-th prompt. ,in, The larger the value, the higher the novelty of the corresponding candidate sequence relative to other candidate sequences within the same candidate group. The collision rate or repetition rate of the corresponding candidate sequence is defined as:

[0149] in, This represents the collision rate or repetition rate of the j-th candidate sequence corresponding to the i-th cue. As the diversity score of a candidate sequence increases, its collision rate or repetition rate decreases accordingly; conversely, as the repetition of a candidate sequence with other candidate sequences within the same candidate group increases, its diversity score decreases accordingly. The diversity score is used to characterize the uniqueness of each candidate sequence within the group during the candidate sequence sampling phase and serves as a modulation factor for the basic advantage, without replacing the original task reward calculation, intra-group reward normalization, and global batch normalization processes in the GDPO algorithm. During the candidate sequence unpacking and sampling phase, the DMA algorithm scales the GDPO basic advantage based on the diversity score. The modulation advantage of the j-th candidate sequence corresponding to the i-th cue. Defined as:

[0150] in, These are hyperparameters used to control the intensity of diverse modulations. For symbolic functions. When the fundamental advantage When the base advantage is positive, the sign function is 1, and the diversity score is used to increase the magnitude of the positive advantage; when the base advantage is negative, the sign function is -1, and the diversity score is used to adjust the magnitude of the negative advantage; when the base advantage is 0, the modulated advantage remains 0. By incorporating the sign of the advantage and the diversity score of the candidate sequence into the scaling factor, different policy update weights can be obtained for candidate sequences with different diversity levels while preserving the direction of the base advantage. Expanding the above modulation formula, the modulation advantage can be expressed in the following piecewise form:

[0151] When the base advantage is positive, a higher diversity score results in a larger modulated positive advantage, thus increasing the contribution of novel positive reward candidate sequences to policy updates; when the base advantage is negative, the diversity score, through a coefficient... Adjust the magnitude of the negative advantage; when the base advantage is zero, no corresponding policy update signal is generated. Modulated advantage. This is used to replace the base advantage that has not been modulated by diversity in subsequent policy loss calculations. In this way, the DMA algorithm, without changing the original reward model and reward function definition, directly applies the novelty information of candidate sequences to the advantage magnitude, thereby adjusting the relative contribution of different candidate sequences to the policy model parameter updates.

[0152] 1.4 GDPO-DynClip with Novelty-Aware Dynamic Clipping In one implementation, based on or independent of the diversity modulation advantage method, the pruning boundary in the PPO policy update process is dynamically adjusted according to the novelty of the candidate sequences, forming a novelty-aware dynamic pruning GDPO algorithm, namely the Dynamic Clip algorithm. Unlike applying the same static pruning threshold to all candidate sequences in a training batch, Dynamic Clip calculates local pruning parameters for each candidate sequence, making the policy update magnitude related to the diversity or repetition of that candidate sequence within its candidate group. For the i-th cue in the training batch and the j-th candidate sequence generated based on that cue, the local pruning parameters are... During the optimization phase, the dynamic calculation is as follows:

[0153] in, Indicates the baseline clipping threshold. This represents the attenuation coefficient used to control the sensitivity of the pruning range to the repetition rate of the candidate sequence. This represents the diversity score of the j-th candidate sequence corresponding to the i-th prompt. This represents the collision rate or repetition rate of the corresponding candidate sequence. When a candidate sequence has a high diversity score, the exponential decay decreases, and the local pruning parameter approaches the baseline pruning threshold; when a candidate sequence has a high repetition rate, the exponential decay increases, and the local pruning parameter decreases accordingly, but not below the minimum allowed pruning value. After applying the local pruning parameter, the strategy probability of the j-th candidate sequence corresponding to the i-th prompt is compared to... Defined as:

[0154] in, Indicates the current policy model. This represents the old strategy model used when generating the current candidate sequence. This indicates the i-th prompt input. This represents the j-th candidate sequence generated for the i-th prompt. The policy probability ratio is used to represent the degree of change in the probability of the current policy model generating the candidate sequence relative to the old policy model. By setting different local pruning parameters for different candidate sequences, the pruning range can vary with the novelty of the candidate sequence, instead of being uniformly limited to all samples by a single fixed pruning threshold. Based on the above policy probability ratio and local pruning parameters, the pruning proxy loss function corresponding to Dynamic Clip is... Defined as:

[0155] in, This represents the set of prompts in the current training batch. G represents the number of prompts in the training batch, and G represents the number of candidate sequences corresponding to each prompt. This represents the advantage used to calculate the policy loss of the j-th candidate sequence corresponding to the i-th cue. Advantages that can be obtained from the standard GDPO algorithm The advantages of multi-modulation can also be utilized when DMA and DynamicClip are used in combination. The loss function calculates the product of the unpruned policy probability ratio and the advantage, and the product of the policy probability ratio and the advantage after applying local pruning boundary constraints, and takes the smaller of the two values ​​for batch averaging during training.

[0156] 2. ESMFold-based de novo binding protein design and antibody structure design reward plugin In one implementation, during reinforcement learning training, the reward function assembler agent can invoke an ESMFold-based structure reward plugin as a structure proxy suitable for designing receptor-binding protein complexes. This structure reward plugin is used to efficiently calculate structure-related rewards during protein sequence optimization, while the final benchmark evaluation of the binding protein is performed separately using AlphaFold-Multimer. Thus, ESMFold serves as an internal reward proxy during reinforcement learning, and AlphaFold-Multimer serves as the final structure evaluation tool after candidate sequence generation. For each receptor-binding protein sequence, the receptor and binding protein sequences are linked by a flexible glycine linker peptide, and the resulting fusion sequence is input into ESMFold. The structure of the fusion sequence is predicted through a single forward propagation. The flexible linker peptide is only used to allow the receptor and binding protein sequences to be input into ESMFold as a single fusion sequence and does not participate in structure reward calculation; when calculating the structure evaluation index, the residues corresponding to the flexible linker peptide are excluded from the set of valid residues. Let V represent the set of all valid residues in the receptor and binding protein sequences. ,in, Indicates the effective number of residues in the receptor sequence. This represents the number of effective residues in the binding protein sequence. The structure reward plugin calculates residue-by-residue structure confidence, global structure confidence, and inter-chain interface confidence based on the output of ESMFold. The predicted local distance difference test score (pLDDT) is used to characterize residue-by-residue structure confidence. The structure reward plugin averages the pLDDT corresponding to all effective residues in the receptor sequence and the binding protein sequence to obtain the average pLDDT of the complex sequence.

[0157] in, The predicted local distance difference test score (pLDDT) represents the score corresponding to the i-th residue in the effective residue set V, where N represents the total number of effective residues in the receptor and binding protein sequences. Residues corresponding to flexible linker peptides do not belong to the effective residue set V and are therefore not included in the average pLDDT calculation. The average pLDDT can be used as a residue-by-residue structure quality reward in the reinforcement learning reward function to evaluate the local structure confidence when the candidate binding protein forms a predicted complex with the receptor. The prediction template modeling score (pTM) is calculated based on the prediction alignment error matrix P output by ESMFold. First, the distance scale is determined according to the total number of effective residues N. :

[0158] Subsequently, based on the predicted alignment error between the i-th residue and the j-th residue in the predicted alignment error matrix... Calculate the corresponding template modeling matrix elements :

[0159] in, This represents the predicted alignment error between residue i and residue j in the predicted alignment error matrix P. This represents the distance scale determined based on the total number of valid residues. This represents the template modeling score corresponding to residue i and residue j. As the prediction alignment error increases... Increase The corresponding decrease; as the prediction alignment error decreases. Decrease The corresponding increase is observed. The template modeling matrix is ​​used to further calculate the global structural confidence and receptor-binding protein interchain interface confidence of the fusion sequence. The sequence-level pTM score is calculated using the maximum value of the average template modeling score of each row in the effective residue set:

[0160] Where i and j represent residue indices in the effective residue set V, and N represents the total number of effective residues. Indicates the prediction alignment error The calculated template modeling score (pTM) is used to characterize the global folding confidence of the predicted structure formed by the receptor sequence and the binding protein sequence, and can be used as the global structure reward component in the reinforcement learning reward function. The interface prediction template modeling score (ipTM) is used to evaluate the inter-chain confidence between receptor residues and binding protein residues. An interface mask is defined. When the i-th residue and the j-th residue belong to different chains, When the i-th residue and the j-th residue belong to the same chain, The IPTM score is calculated as follows:

[0161] in, Used to screen for residue pairs between the receptor chain and the binding protein chain. This represents the template modeling score for the corresponding residue pair. Used to avoid a denominator of zero.

[0162] 3. Statistical Analysis Methods In one implementation, unless otherwise stated, all quantitative data are expressed as mean ± standard deviation. For comparisons at the sequence level or design outcome level, the Mann-Whitney U test is used for statistical analysis because the evaluated metrics often have limited ranges, skewed distributions, or are not pre-assumed to follow a normal distribution. In benchmark comparisons of different methods, a two-tailed test is used by default to avoid pre-assuming the direction of differences between methods. A one-tailed test is used only when analyzing variants of the GDPO algorithm and directional assumptions are pre-defined, such as whether the corresponding algorithm variant improves the reward model score or the diversity of candidate sequences relative to GDPO. For sequence repetition ratios obtained with independent runs as the statistical unit, Welch's t test is used for comparison without assuming equal variances. For binary pass / fail results of the therapeutic antibody profiling TAP, a chi-square test is used for comparison based on the number of pass and fail candidates in each group. For evaluations involving multiple statistical comparisons, appropriate multiple comparison correction methods are used according to the comparison objects and the pre-defined number of comparisons. When the comparison set is small and the pairings to be compared are predetermined, Bonferroni correction is used to control for Type I errors in multiple comparisons. When benchmark method comparisons are performed on multiple targets or multiple evaluation indicators, Benjamini-Hochberg false discovery rate correction is used to control for false discovery rates in multiple tests. Unless otherwise stated, a corrected p-value less than 0.05 indicates a statistically significant difference, where * indicates p < 0.05, ** indicates p < 0.01, and *** indicates p < 0.001. For linear associations between two continuous variables, the Pearson correlation coefficient r is used; for rank relationships between variables, or in cases where the performance of a ranking model is evaluated, the Spearman rank correlation coefficient is used where applicable. The statistical analysis methods are used to evaluate the differences or correlations among different algorithms, candidate sets, or design methods in terms of reward model scores, sequence diversity, structural confidence, functional prediction values, exploitability indicators, and other quantitative evaluation indicators.

[0163] 4. UMAP Analysis Method In one implementation, the Unified Manifold Approximation and Projection Method (UMAP) is used to reduce the dimensionality and visualize the distribution of protein sequences in a high-dimensional sequence space. First, the generated and reference sequences to be analyzed undergo amino acid validity checks, sequence format standardization, and necessary length adjustments. Then, one-hot encoding is used to convert each protein sequence into a numerical representation. For each sequence position, a one-hot vector corresponding to the amino acid category represents the residue type at that position, and the one-hot vectors of each position are combined to form a high-dimensional feature representation of the corresponding protein sequence. Subsequently, the UMAP algorithm implemented using UMAP-learn projects these high-dimensional features into a two-dimensional space. UMAP first estimates the local neighborhood relationships between sequences in the original high-dimensional sequence space and represents these relationships as a weighted fuzzy graph. Then, it constructs a low-dimensional graph structure corresponding to the high-dimensional fuzzy graph in the low-dimensional space and optimizes the two-dimensional embedding coordinates using stochastic gradient descent to minimize the cross-entropy between the high-dimensional and low-dimensional graph representations. The two-dimensional coordinates obtained after UMAP dimensionality reduction are used to represent the relative positions of each protein sequence in the low-dimensional space, and to observe the spatial organization, clustering state, distribution range, and overlap between different candidate sets of the generated and reference sequences. Each two-dimensional data point corresponds to a protein sequence, and can be marked with different colors or symbols according to the generation method, training stage, candidate type, reward interval, or other preset grouping information. When comparing different reinforcement learning algorithms, each algorithm uses the same sequence encoding method, UMAP parameters, and random seed to ensure the comparability of the two-dimensional distribution results of different candidate sets. The two-dimensional coordinates are only used to characterize the relative distribution and clustering relationship of high-dimensional sequence features and are not directly used as evaluation results for protein function, structural quality, or binding ability; the functional and structural performance of candidate sequences are still independently determined by the corresponding reward model, structural prediction model, or experimental evaluation indicators.

[0164] 5. Minimum Spanning Tree Analysis Method In one implementation, a minimum spanning tree is used to characterize the spatial topology of candidate sequences generated during reinforcement learning training. First, the Levenshtein edit distance between any two candidate protein sequences is calculated, and a distance graph of the candidate sequences is constructed, using each candidate sequence as a node and the pairwise Levenshtein distance between candidate sequences as edge weights. A minimum spanning tree is then calculated on this distance graph to connect all candidate sequence nodes while minimizing the total weight of the connecting edges. The total weight of the minimum spanning tree is used to quantify the cumulative coverage of candidate sequences in the sequence space; the total weight is the sum of the weights of all connecting edges in the minimum spanning tree. The diameter of the minimum spanning tree is used to quantify the compactness of the explored sequence regions; the diameter is the maximum path length between any two nodes in the minimum spanning tree. Candidate sequence nodes can also be labeled according to their corresponding training iterations to characterize the position, connectivity, and expansion path of candidate sequences generated at different training stages in the minimum spanning tree. In one implementation, the number of non-repeating candidate sequences accumulated during reinforcement learning training is statistically analyzed, and an exponential saturation model is used to fit the change in the cumulative number of non-repeating sequences with training time or iteration count.

[0165] Where N(t) represents the cumulative number of non-repeating candidate sequences obtained before training time or iteration position t, C represents the asymptotic capacity of the cumulative number of non-repeating candidate sequences, and k represents the instantaneous discovery rate of candidate sequences. As t increases, N(t) gradually approaches the asymptotic capacity C. Using the total weight of the minimum spanning tree, the diameter of the minimum spanning tree, the asymptotic capacity C, and the instantaneous discovery rate k, the sequence space exploration process of different reinforcement learning algorithms can be characterized in terms of the cumulative coverage of the candidate sequence space, the compactness of the space, the potential sequence discovery capacity, and the sequence discovery speed.

[0166] 6. TAP Analysis Method In one implementation, for each pair of paired VH-VL candidate sequences, ABodyBuilder2 is used to predict the three-dimensional structure of the corresponding antibody Fv fragment. Since NACCESS is not available in high-throughput analysis environments, the therapeutic antibody profiling metrics are recalculated using FreeSASA as the solvent-accessible surface area calculation tool (Mitternacht, F1000Research 2016, 5, 189.) according to the five therapeutic antibody profiling metrics defined by Raybould et al. (Raybould et al., Proc. Natl. Acad. Sci. USA 2019, 116, 4025-4030.). These five metrics include the total length of the complementarity-determining region (CD), the surface hydrophobic plaques (PSH) near the CD, the surface positively charged plaques (PPC) near the CD, the surface negatively charged plaques (PNC) near the CD, and the charge symmetry parameter SFvCSP of the antibody Fv fragment. The total length of the complementarity-determining region (PDG) is defined as the sum of the number of amino acid residues in the six PRGs; PSH is defined as the largest continuous hydrophobic surface patch near the PRG; PPC and PNC are defined as the largest aggregated positively charged patch and the largest aggregated negatively charged patch near the PRG, respectively. The charge symmetry parameter SFvCSP of the antibody Fv fragment is defined as:

[0167] in, and The net surface charges of the heavy and light chain variable regions of the antibody are represented, respectively, and r represents the centroid distance between the heavy and light chain variable regions. SFvCSP is used to characterize the symmetry relationship between the surface charges of the heavy and light chain variable regions. The net surface charges of the heavy and light chain variable regions are calculated based on their three-dimensional structure and solvent-accessible surface area, and the centroid distance between the domains is determined based on the three-dimensional coordinates of the heavy and light chain variable regions. Since the solvent-accessible surface area values ​​calculated by FreeSASA and NACCESS are not completely consistent, the absolute risk threshold in the original therapeutic antibody profiling analysis is not directly used. Instead, the same FreeSASA analysis workflow is used to recalculate all five TAP indicators for the clinical-stage therapeutic antibody reference set, and the green, yellow, and red risk boundaries are redefined according to the original percentile strategy. For the total length of the complementarity-determining region (PDG) and PSH, a bilateral boundary was used. The middle 90% range between the 5th and 95th percentiles of the clinical-stage therapeutic antibody reference set distribution was designated green; values ​​within the measured range of the reference set and located in the lowest or highest 5% were designated yellow; and values ​​below or above the measured range of the reference set were designated red. For PPC and PNC, an upper tail boundary was used. Values ​​below the 95th percentile were designated green, values ​​within the highest 5% of the measured range of the clinical-stage therapeutic antibody reference set were designated yellow, and values ​​above the measured range of the reference set were designated red. For SFvCSP, a lower tail boundary was used. Values ​​above the 5th percentile were designated green, values ​​within the lowest 5% of the measured range of the clinical-stage therapeutic antibody reference set were designated yellow, and values ​​below the measured range of the reference set were designated red. The five TAP parameters for both reference and candidate antibodies were recalculated using the same FreeSASA workflow to ensure that the risk stratification of candidate antibodies matched the solvent accessible surface area calculation method. In one implementation, the TAP pass rate is defined as the proportion of candidate antibodies whose five therapeutic antibody profiling metrics are all classified as green out of the total number of candidate antibodies. If the total complementarity-determining region length, PSH, PPC, PNC, and SFvCSP of a candidate antibody are all in the green range, the candidate antibody is judged to have passed the TAP; if at least one metric is classified as yellow or red, the candidate antibody is judged to have failed the TAP. The system can write the raw values ​​of the five metrics, color classification, reasons for failure, and overall TAP pass status into a candidate evaluation table, and use the TAP pass rate to compare the exploitability characteristics of different candidate sets or different antibody design methods.

Claims

1. A protein sequence computation and experimental closed-loop design system based on candidate group diversity regulation, characterized in that, This includes a cloud computing and supercomputing platform, a protein sequence design intelligent agent subsystem, a biological verification laboratory subsystem, and an experimental feedback interface; The cloud computing and supercomputing platform is used to provide computing and data resources for protein sequence generation, model training, reinforcement learning optimization, structure prediction, and candidate evaluation; the protein sequence design intelligent agent subsystem includes a task input module, a protein sequence generator, a reward calculation module, a reinforcement learning optimization module, and a candidate sequence output module; The task input module is used to receive a protein design task, and at least one of a protein sequence, target sequence, antigen sequence, experimental data, or design constraints corresponding to the protein design task; the protein sequence generator is used to generate a candidate group for the same task input, the candidate group including multiple candidate protein sequences generated in the same sampling round or optimization round; the reward calculation module is used to calculate one or more reward components corresponding to each candidate protein sequence, and normalize each reward component within the candidate group respectively; The reinforcement learning optimization module is used to determine the advantage of each candidate protein sequence based on the normalized reward components, adjust the advantage or policy update according to the repetition, diversity, or word-level policy probability distribution differences among candidate protein sequences within the candidate group, and update the protein sequence generator based on the adjusted advantage or policy update method; wherein, the method of adjusting the advantage or policy update includes at least one of the following: determining the group-level diversity reward based on the repetition of the candidate protein sequence within the candidate group, and combining the advantage corresponding to the group-level diversity reward with the advantages corresponding to other reward components; scaling the advantage according to the diversity score of the candidate protein sequence and the positive or negative direction of its advantage; dynamically adjusting the pruning threshold of the policy probability ratio according to the repetition of the candidate protein sequence; and adjusting the strategy probability ratio based on the different candidate protein sequences within the candidate group. The divergence between the probability distributions of word-level strategies in white matter sequences is used to construct an exclusionary regularization term; the candidate sequence output module is used to output the candidate protein sequences generated by the updated protein sequence generator, along with corresponding candidate numbers and experimental task information; the biological validation laboratory subsystem includes at least a candidate protein preparation unit and a property detection unit, wherein the candidate protein preparation unit is used to prepare corresponding candidate protein experimental samples based on the candidate protein sequences, and the property detection unit is used to detect the candidate protein experimental samples and obtain at least one experimental measurement result; the experimental feedback interface is used to associate the experimental measurement results with the candidate protein sequences based on the candidate numbers, forming structured feedback data, and inputting the structured feedback data into the protein sequence design intelligent agent subsystem for use in the next round of protein sequence generation or reinforcement learning optimization.

2. The protein sequence calculation and experimental closed-loop design system according to claim 1, characterized in that, The reinforcement learning optimization module employs a group-based relative policy optimization method based on exclusivity divergence regularization, including: calculating the Kullback-Leibler divergence of the term-level policy probability distribution of the current policy model relative to the corresponding term-level policy probability distribution of the reference model; the Kullback-Leibler divergence, abbreviated as KL divergence, is used to constrain the policy shift of the current policy model relative to the reference model; and calculating the Jensen-Shannon divergence between the term-level policy probability distributions of different candidate protein sequences within the same candidate group at common effective term positions. Jensen-Shannon divergence, or JS divergence for short, is used to characterize the differences between the term-level policy probability distributions of different candidate protein sequences. A candidate group-level exclusionary divergence is determined based on the JS divergence. A joint objective function is constructed based on the pruning proxy policy loss, the KL divergence, and the candidate group-level exclusionary divergence. The KL divergence is introduced into the joint objective function with positive weights as a regularization term to limit the deviation of the current policy model from the reference model, and the candidate group-level exclusionary divergence is introduced into the joint objective function with negative weights as a regularization term to promote differences in policy probability distributions within candidate groups.

3. The protein sequence calculation and experimental closed-loop design system according to claim 2, characterized in that, The reinforcement learning optimization module is used to normalize each reward component in the candidate group, combine the normalized reward components into a total advantage according to the weights corresponding to each reward component, and perform a second normalization on the total advantage in the training batch containing multiple candidate groups to obtain the final advantage used to calculate the loss of the pruning agent strategy.

4. The protein sequence calculation and experimental closed-loop design system according to claim 3, characterized in that, The reinforcement learning optimization module is used to calculate the ratio of the conditional probability of each effective word in the candidate protein sequence generated by the current policy model to the conditional probability of the corresponding effective word generated by the old policy model, so as to obtain the policy probability ratio; calculate the product of the policy probability ratio and the final advantage, and the product of the policy probability ratio after being restricted by a preset pruning interval and the final advantage; determine the pruning proxy policy loss based on the smaller value of the two products.

5. The protein sequence calculation and experimental closed-loop design system according to claim 2, characterized in that, The reinforcement learning optimization module uses a low-variance estimation method to approximate the KL divergence, including: recording the difference between the log probability of the reference model for the corresponding effective word and the log probability of the current policy model for the effective word as the log probability difference; subtracting the log probability difference from the exponential value of the log probability difference and subtracting 1 to obtain the KL divergence estimate of the corresponding effective word; and averaging the KL divergence estimates of only the non-filled effective words in the candidate protein sequence based on the indication information of the non-filled effective words to obtain the sequence-level average KL divergence.

6. The protein sequence calculation and experimental closed-loop design system according to claim 2, characterized in that, The terminology refers to the basic sequence unit used by the protein sequence generator when performing sequence encoding, generation, or probability calculation. The basic sequence unit includes at least one of amino acids, inter-strand separators, sequence start symbols, or sequence stop symbols. The effective terminology refers to the terminology in the candidate protein sequence that participates in the policy loss or divergence calculation, excluding padding symbols. For the i-th and k-th candidate protein sequences within the same candidate group, the reinforcement learning optimization module is used to determine the common effective terminology positions of the two candidate protein sequences and calculate the JS divergence between the terminology-level policy probability distributions of the two candidate protein sequences only at the common effective terminology positions.

7. The protein sequence calculation and experimental closed-loop design system according to claim 6, characterized in that, The reinforcement learning optimization module calculates the average distribution of the term-level policy probability distribution of the i-th candidate protein sequence and the term-level policy probability distribution of the k-th candidate protein sequence at each common effective term position to obtain a mixed probability distribution; it then calculates the KL divergence of the term-level policy probability distribution of the two candidate protein sequences relative to the mixed probability distribution; and determines the average of the two KL divergences as the term-level JS divergence at the corresponding effective term position.

8. The protein sequence calculation and experimental closed-loop design system according to claim 7, characterized in that, The reinforcement learning optimization module averages the lexical-level JS divergence at all valid lexical positions shared by the i-th and k-th candidate protein sequences to obtain the sequence-level JS divergence between the two candidate protein sequences; it then traverses each unordered and non-repeating candidate sequence pair within the same candidate group, calculates the sequence-level JS divergence of each candidate sequence pair, and averages the sequence-level JS divergence of each candidate sequence pair to obtain the candidate group-level exclusion divergence.

9. The protein sequence calculation and experimental closed-loop design system according to claim 2, characterized in that, The reward components include at least two of the following: functional prediction reward, sequence naturalness reward, amino acid legitimacy reward, length constraint reward, structure confidence reward, binding interface reward, fixed pairing reward, antibody heavy chain and light chain pairing reward, diversity reward, and exploitability reward; the protein design task includes at least one of the following: data-driven protein redesign, de novo binding protein design, de novo antibody design, and experimental data-guided protein sequence optimization.

10. The protein sequence calculation and experimental closed-loop design system according to claim 2, characterized in that, It also includes a structure reward plugin based on a protein language model; the structure reward plugin is used to link the receptor sequence and the candidate binding protein sequence into a fusion sequence through a flexible linker peptide, and use the ESMFold structure prediction model to predict the structure of the fusion sequence. After excluding the residues corresponding to the flexible linker peptide, at least one of the following indicators is calculated based on the effective residues of the receptor sequence and the candidate binding protein sequence: the predicted local distance difference test score pLDDT for characterizing residue-by-residue structure confidence, the predicted template modeling score pTM for characterizing global folding confidence, and the interface structure confidence score determined based on the prediction alignment error matrix output by the structure prediction model and the position mask between the receptor residues and the candidate binding protein residues. And a structural reward is formed in the reinforcement learning process based on at least one of the aforementioned metrics.

11. A protein sequence computation and experimental closed-loop evolutionary design system based on a self-evolving multi-agent system, characterized in that, This includes a cloud computing and supercomputing platform, a self-evolving multi-agent computational design subsystem, a biological verification laboratory subsystem, and an experimental feedback interface; The cloud computing and supercomputing platform is used to provide computing and data resources for model training, protein sequence generation, reinforcement learning, structure prediction, molecular docking, and candidate evaluation; the self-evolving multi-agent computational design subsystem includes a task input module, an orchestrator, an evaluator, multiple specialized sub-agents, a product registration module, a memory system, and a candidate sequence output interface. The task input module is used to receive protein design tasks, as well as protein sequence data, target sequence data, antigen sequence data, experimental data or design constraints corresponding to the protein design tasks. The orchestrator is used to generate a workflow containing multiple typified task items based on the protein design task and historical design information in the memory system, and to schedule the multiple specialized sub-agents; the evaluator is used to review the workflow, structured input, intermediate products, and candidate protein sequences before and after the execution of the typified task items, and to prevent the corresponding typified task items from continuing to execute when the review results do not meet preset conditions, generate correction information, and cause the orchestrator to replan the workflow; the multiple specialized sub-agents include at least a task construction sub-agent, a data processing sub-agent, a generator task adaptation sub-agent, a reward model training and evaluation sub-agent, and a reward function assembly sub-agent. The system comprises a primary agent, a reinforcement learning sub-agent, and an evaluation sub-agent. The reinforcement learning sub-agent is used to normalize multiple reward components within a candidate group, perform secondary normalization on the total advantage obtained based on the multiple reward components within a training batch containing multiple candidate groups, and update the protein sequence generator based on the pruned agent policy loss, the Kullback-Leibler divergence between the current policy model and the reference model, and the candidate group-level repulsion Jensen-Shannon divergence between different candidate protein sequences within the same candidate group. The Kullback-Leibler divergence is used to constrain the policy shift of the current policy model relative to the reference model, and the candidate group-level repulsion Jensen-Shannon divergence... Hannon divergence is used to facilitate differences in the probability distributions of lexical-level strategies among different candidate protein sequences within a candidate group; the product registration module is used to register generator products, reward models, and reward functions that pass preset admission conditions, and to restrict downstream dedicated sub-agents from calling products that are not registered or have not passed preset admission conditions; the memory system is used to save task status, workflow, intermediate products, candidate protein sequences, evaluation results, experimental feedback data, failure reasons, and correction information, and to use the saved information for the next round of workflow generation and candidate protein sequence optimization; the candidate sequence output interface is used to output candidate protein sequences, candidate numbers, and experimental task information to the biological verification laboratory subsystem; the biological verification laboratory The subsystem includes at least a candidate protein preparation unit and a property detection unit. The candidate protein preparation unit is used to prepare corresponding candidate protein experimental samples according to the candidate protein sequence. The property detection unit is used to detect the candidate protein experimental samples according to the experimental task information and obtain experimental measurement results. The experimental feedback interface is used to associate the experimental measurement results with the candidate protein sequence, generation round, and evaluation results according to the candidate number, convert them into structured feedback data, and input the structured feedback data into at least one of the following: data processing sub-agent, reward model training and evaluation sub-agent, reward function assembly sub-agent, evaluation sub-agent, or memory system, for use in the next round of workflow generation and candidate protein sequence optimization.

12. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 11, characterized in that, Each of the typified task items includes at least a task type, task instructions, structured input, expected output, task reason, resource type, and dependencies; the task type includes at least one of task construction, data processing, task adaptive pre-training, supervised fine-tuning, reward model training and evaluation, reward function assembly, reinforcement learning evolution, candidate evaluation, and task termination.

13. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 11, characterized in that, The evaluator includes a rule review function and a semantic review function; the rule review function is used to check task dependencies, input and output formats, data paths, product registration status, product quality status, reward function test status, and execution conditions; the semantic review function is used to determine whether the selected dedicated sub-agent, data field role, reward target, and candidate protein sequence are consistent with the protein design task.

14. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 11, characterized in that, The product registration module includes a generator product registry, a reward model registry, and a reward function registry. The generator products generated by the generator task adaptation sub-agent are written to the generator product registry. Reward models that achieve the corresponding quality threshold on the validation or test set based on preset evaluation metrics are written to the reward model registry. Reward functions that complete reward intent review, input / output field checks, and procedural testing are written to the reward function registry. The reward function assembly sub-agent only calls registered reward models that have reached the corresponding quality threshold, and the reinforcement learning sub-agent only calls registered generator products and reward functions.

15. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 11, characterized in that, The memory system includes a short-term memory unit, a memory compressor, an insight memory unit, and a long-term memory unit. The short-term memory unit is used to store the workflow, execution record, candidate protein sequence, and evaluation results of the current task. The memory compressor is used to extract the task type, execution status, evaluation index, failure reason, and correction information from the short-term memory unit when the record volume of the current task reaches a preset context capacity condition, and generate a compressed record. The insight memory unit is used to store filtered compressed records; The long-term memory unit is used to store cross-task historical information and is available for retrieval by the orchestrator and evaluator.

16. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 11, characterized in that, The cloud computing and supercomputing platform includes a communication unit, a storage unit, a computing unit, an input / output interface, a task scheduling unit, and an execution sandbox; the computing unit includes at least one of a central processing unit (CPU) computing resources and a graphics processing unit (GPU) computing resources, used to perform model training, reinforcement learning, structure prediction, molecular docking, and candidate evaluation; the storage unit is used to store input data, model products, reward models, reward functions, candidate protein sequences, and structured feedback data. The task scheduling unit is used to allocate computing resources according to task resource type and dependency relationship; The execution sandbox is used to execute the training programs, evaluation programs, or external tools called by the multiple dedicated sub-agents in a runtime environment isolated from the orchestrator main process.

17. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 11, characterized in that, The data processing sub-agent is used to identify at least one of the protein sequence field, tag field, receptor sequence field, binding protein sequence field, antibody heavy chain sequence field, and antibody light chain sequence field in the input data, and to perform at least one of the following on the input data: cleaning, deduplication, amino acid validity check, sequence length statistics, tag data type conversion, tag dimension conversion, and dataset partitioning.

18. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 14, characterized in that, The generator task adaptor is used to adapt the basic protein sequence generator through at least one of task adaptive pre-training and supervised fine-tuning, and writes at least one of the following as generator products into the generator product registry: model configuration, vocabulary, initialized model parameters, post-trained model parameters, training input template, chain format and sequence length statistics.

19. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 11, characterized in that, The reward model training and evaluation sub-agent is used to train a single-chain reward model based on protein sequences and corresponding functional tags, or to train a multi-chain reward model based on receptor sequences, binding protein sequences and corresponding tags. It determines whether the reward model meets the quality admission conditions based on whether at least one preset evaluation index of the reward model on the validation set or test set reaches the corresponding quality threshold.

20. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 19, characterized in that, When the reward model obtained by the reward model training and evaluation sub-agent does not meet the quality admission conditions, the evaluator prevents the reward model from entering the reward function assembly process. The orchestrator causes the reward function assembly sub-agent to construct an alternative reward function by using at least one of the following: structure prediction score, molecular docking score, or sequence distance or distribution distance between the candidate protein sequence and the high-quality experimental samples determined by experimental indicators or experimental ranking.

21. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 11, characterized in that, The evaluation sub-agent is used to perform at least one of the following on candidate protein sequences: functional prediction, sequence naturalness evaluation, global structural confidence evaluation, residue-by-residue structural confidence evaluation, binding interface evaluation, sequence novelty evaluation, and exploitability evaluation. Based on the evaluation results, it filters, sorts, or determines the experimental verification priority of candidate protein sequences to form an experimental candidate list output to the biological verification laboratory subsystem.

22. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 11, characterized in that, The candidate protein preparation unit includes at least one of a sequence synthesis unit, a vector construction unit, a protein expression unit, and a protein purification unit; the property detection unit includes at least one of an expression level detection unit, a purification yield detection unit, a functional activity detection unit, a binding detection unit, an affinity detection unit, a stability detection unit, an aggregation detection unit, a cell binding detection unit, an immunogenicity-related detection unit, and an exploitability detection unit; the experimental feedback interface includes an experimental task output interface, an experimental result receiving interface, and a feedback data conversion unit. The experimental task output interface is used to send an experimental candidate list and experimental task information to the biological validation laboratory subsystem. The experimental result receiving interface is used to receive experimental measurement results. The feedback data conversion unit is used to associate the experimental measurement results with the candidate protein sequence, generation round, evaluation result, and experimental batch according to the candidate number, and form structured feedback data.

23. The protein sequence calculation and experimental closed-loop evolutionary design system according to claim 11, characterized in that, The protein design task includes at least one of data-driven protein redesign, de novo binding protein design, de novo antibody design, and experimental data-guided protein sequence optimization. For data-driven protein redesign tasks, the property detection unit is used to obtain at least one of the following: expression level, functional activity, stability, or solubility of the candidate protein; for de novo binding protein design tasks, the property detection unit is used to obtain at least one of the following: binding signal, binding affinity, functional activity, stability, or immunogenicity-related indicators of the candidate binding protein; for de novo antibody design or experimental data-guided antibody optimization tasks, the property detection unit is used to obtain at least one of the following: binding affinity, selectivity, stability, aggregation tendency, or exploitability indicators of the candidate antibody.

24. A method for protein sequence calculation and experimental closed-loop design based on candidate group exclusion divergence regularization, characterized in that, include: Receive a protein design task, and at least one of the following: protein sequence, target sequence, antigen sequence, experimental data, or design constraints corresponding to the protein design task; By invoking the strategy model through cloud computing and supercomputing platforms, candidate groups containing multiple candidate protein sequences are generated for the same task input; Calculate multiple reward components corresponding to each candidate protein sequence; Each reward component is normalized within the candidate group, and the normalized reward components are combined into a total advantage according to the weights corresponding to each reward component. The total advantage is normalized a second time within a training batch containing multiple candidate groups. The ratio of the conditional probability of the current strategy model generating each effective word to the conditional probability of the old strategy model generating the corresponding effective word is calculated to obtain the strategy probability ratio. The pruning proxy strategy loss is calculated based on the total advantage after the second normalization, the strategy probability ratio, and the preset pruning threshold. Calculate the Kullback-Leibler divergence of the lexical-level policy probability distribution of the current policy model relative to the corresponding lexical-level policy probability distribution of the reference model. The Kullback-Leibler divergence is used to constrain the policy shift of the current policy model relative to the reference model. For different candidate protein sequences within the same candidate group, calculate the Jensen-Shannon divergence between the lexical-level policy probability distributions of the two candidate protein sequences at common effective lexical positions. The Jensen-Shannon divergence is used to characterize the difference between the lexical-level policy probability distributions of the two candidate protein sequences. Determine the sequence-level repulsion divergence based on the Jensen-Shannon divergence at each common effective lexical position, and determine the candidate group-level repulsion divergence based on the sequence-level repulsion divergence of all non-repeating candidate sequence pairs within the same candidate group. Based on the pruning proxy policy loss, A joint objective function is constructed using Kullback-Leibler divergence and candidate group-level exclusionary divergence, wherein the Kullback-Leibler divergence is introduced into the joint objective function with positive weights, and the candidate group-level exclusionary divergence is introduced into the joint objective function with negative weights. The policy model is updated by minimizing the joint objective function. Candidate protein sequences, corresponding candidate numbers, and experimental task information are generated using the updated policy model. The candidate protein sequences and experimental task information are sent to the biological validation laboratory subsystem, where the candidate protein preparation unit prepares candidate protein experimental samples based on the candidate protein sequences, and the property detection unit obtains the experimental measurement results of the candidate protein experimental samples. The experimental measurement results are associated with the candidate protein sequences according to the candidate numbers and converted into structured feedback data. The structured feedback data is input into the next round of reward calculation or policy model update.

25. The protein sequence calculation and experimental closed-loop design method according to claim 24, characterized in that, The calculation of the candidate group-level exclusionary divergence includes: for any two candidate protein sequences within the same candidate group, determining the common effective term positions of the two candidate protein sequences; averaging the term-level strategy probability distributions of the two candidate protein sequences at each common effective term position to obtain a mixed probability distribution; calculating the Kullback-Leibler divergence of the term-level strategy probability distributions of the two candidate protein sequences relative to the mixed probability distribution; determining the average of the two Kullback-Leibler divergences as the Jensen-Shannon divergence at the corresponding effective term position; averaging the Jensen-Shannon divergences at all common effective term positions to obtain the sequence-level exclusionary divergence; and averaging the sequence-level exclusionary divergences of all disordered and non-repeating candidate sequence pairs within the same candidate group to obtain the candidate group-level exclusionary divergence.

26. The protein sequence calculation and experimental closed-loop design method according to claim 24, characterized in that, Calculating the Kullback-Leibler divergence between the current strategy model and the reference model includes: recording the difference between the log probability of the reference model for the corresponding effective word and the log probability of the current strategy model for the effective word as the log probability difference; subtracting the log probability difference from the exponential value of the log probability difference and then subtracting 1 to obtain the estimated Kullback-Leibler divergence value for the corresponding effective word; and averaging the estimated Kullback-Leibler divergence values ​​for only the non-filled effective words in the candidate protein sequence based on the indication information of the non-filled effective words to obtain the sequence-level average Kullback-Leibler divergence.

27. The protein sequence calculation and experimental closed-loop design method according to claim 24, characterized in that, The multiple reward components include at least two of the following: functional prediction reward, sequence naturalness reward, amino acid legitimacy reward, length constraint reward, structure confidence reward, binding interface reward, fixed pairing reward, antibody heavy chain and light chain pairing reward, diversity reward, and exploitability reward.

28. The protein sequence calculation and experimental closed-loop design method according to claim 24, characterized in that, The protein design task includes at least one of data-driven protein redesign, de novo binding protein design, de novo antibody design, and experimental data-guided protein sequence optimization.

29. The protein sequence calculation and experimental closed-loop design method according to claim 24, characterized in that, Also includes: The receptor sequence and the candidate binding protein sequence are linked together to form a fusion sequence using a flexible linker peptide; The ESMFold structure prediction model is used to predict the structure of the fusion sequence; The residues corresponding to the flexible linker peptide are excluded; at least one of the following is calculated based on the effective residues of the receptor sequence and the candidate binding protein sequence: the predicted local distance difference test score (pLDDT) used to characterize the confidence of the structure per residue and the predicted template modeling score (pTM) used to characterize the confidence of the global folding; the interface structure confidence score is calculated based on the prediction alignment error matrix output by the structure prediction model and the position mask between the receptor residues and the candidate binding protein residues; at least one of the predicted local distance difference test score (pLDDT), the predicted template modeling score (pTM), and the interface structure confidence score is used as the reward component.

30. A method for protein sequence computation and experimental closed-loop evolutionary design based on a self-evolving multi-agent system, characterized in that, include: Receive protein design tasks, as well as protein sequence data, target sequence data, antigen sequence data, experimental data, or design constraints corresponding to the protein design tasks; The orchestrator generates a workflow containing multiple typified task items based on the protein design task and historical design information in the memory system. The evaluator reviews the workflow, structured inputs, task dependencies, and registered products. When the review results meet preset conditions, multiple dedicated sub-agents are scheduled through cloud computing and supercomputing platforms to perform data processing, generator task adaptation, reward model training and evaluation, reward function assembly, reinforcement learning evolution, and structure and extensibility evaluation. The reinforcement learning sub-agent normalizes multiple reward components within each candidate group. Within a training batch containing multiple candidate groups, a second normalization is performed on the total advantage obtained based on these multiple reward components. The protein sequence generator is updated based on the pruned agent policy loss, the Kullback-Leibler divergence between the current policy model and the reference model, and the candidate group-level exclusivity Jensen-Shannon divergence between different candidate protein sequences within the same candidate group. The Kullback-Leibler divergence constrains the policy shift of the current policy model relative to the reference model, and the candidate group-level exclusivity Jensen-Shannon divergence promotes differences in the term-level policy probability distributions of different candidate protein sequences within the candidate group. When the workflow, intermediate products, or candidate protein sequences do not meet preset conditions, the evaluator prevents the corresponding task from continuing and generates correction information. The orchestrator re-plans the workflow based on the correction information. The generator products and rewards that pass the preset admission conditions are then... The incentive model and reward function are written into the corresponding product registry; the task status, workflow, intermediate products, candidate protein sequences, evaluation results, failure reasons, and correction information are written into the memory system; an experimental candidate list and experimental task information are formed based on the evaluation results of the candidate protein sequences, and the experimental candidate list and experimental task information are sent to the biological verification laboratory subsystem through the experimental feedback interface; the candidate protein preparation unit of the biological verification laboratory subsystem prepares candidate protein experimental samples according to the candidate protein sequences, and the property detection unit obtains experimental measurement results according to the experimental task information; the experimental feedback interface associates the experimental measurement results with the candidate protein sequences, generation rounds, and evaluation results according to the candidate numbers, and converts them into structured feedback data; the structured feedback data is input into at least one of the data processing sub-agent, reward model training and evaluation sub-agent, reward function assembly sub-agent, evaluation sub-agent, or memory system; the next round of workflow and candidate sequence optimization strategy are generated based on the structured feedback data and historical information in the memory system.

31. The protein sequence calculation and experimental closed-loop evolutionary design method according to claim 30, characterized in that, The formation and utilization of the structured feedback data includes: receiving at least one of the following scan results corresponding to candidate protein experimental samples: expression, purification, binding, affinity, stability, aggregation, cell binding, functional activity, exploitability, or deep mutation; associating the results with candidate protein sequences, generation rounds, evaluation results, and experimental batches according to the candidate number; converting the associated results into structured feedback data with preset fields; performing at least one of the following processing on the structured feedback data by a data processing sub-agent: field identification, data type conversion, label standardization, quality inspection, and dataset partitioning; training or updating the reward model based on the processed structured feedback data; and writing the reward model into the reward model registry when the reward model meets the quality admission criteria, and using it for the next round of reward function assembly and reinforcement learning evolution.

32. The protein sequence calculation and experimental closed-loop evolutionary design method according to claim 30, characterized in that, When the reward model does not meet the quality admission criteria, the evaluator prevents the reward model from entering the reward function assembly process, and the orchestrator calls the reward function assembly sub-agent to construct an alternative reward function using at least one of the following: structure prediction score, molecular docking score, or sequence distance or distribution distance between the candidate protein sequence and the high-quality experimental samples determined by experimental indicators or experimental ranking.

33. The protein sequence calculation and experimental closed-loop evolutionary design method according to claim 30, characterized in that, The memory system includes a short-term memory unit, a memory compressor, an insight memory unit, and a long-term memory unit. When the information stored in the short-term memory unit reaches a preset context capacity condition, the memory compressor extracts the task type, execution status, evaluation indicators, failure reasons, and correction information from the short-term memory unit and generates a compressed record. The filtered compressed record is written into the insight memory unit or the long-term memory unit and is available for retrieval by the orchestrator and evaluator in subsequent workflows.

34. An electronic device, characterized in that, It includes a processor, a memory, and a communication interface. The memory stores a computer program, which, when executed by the processor, causes the electronic device to perform the computational processing, task scheduling, experimental task information transmission, experimental measurement result reception, and structured feedback data processing steps of the method according to any one of claims 24 to 33.

35. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, causes the electronic device to perform the computational processing, task scheduling, experimental task information transmission, experimental measurement result reception, and structured feedback data processing steps of the method according to any one of claims 24 to 33.

Images