Protein binder iterative optimization method and system based on agent cooperation execution

Abstract

The application provides a protein binder iterative optimization method and system based on agent collaborative execution, relates to the technical field of protein engineering, bioinformatics and artificial intelligence assisted protein design, and can be applied to an agent execution system, the agent execution system comprising multiple functional agents; the entire iterative process can realize model calling, iterative optimization and other functions through the cooperation of multiple agents, so that the optimization efficiency, process interpretability and structure confidence of the protein binder are improved, the calculation cost is reduced, and the generation efficiency of the candidate protein binder is improved.

Description

Technical Field

[0001] This invention relates to the fields of protein engineering, bioinformatics, and artificial intelligence-assisted protein design technology, and in particular to an iterative optimization method and system for protein conjugates based on intelligent agent collaborative execution. Background Technology

[0002] Protein conjugates are polypeptides, antibodies, nanobodies, microproteins, or other protein molecules that can form specific interactions with target proteins. High-affinity and high-specificity protein conjugates have significant applications in drug development, disease diagnosis, targeted delivery, synthetic biology, and molecular recognition. Traditional protein conjugate design typically relies on experimental screening, directed evolution, or rational design methods. While these methods can yield conjugates with good performance, they often require multiple rounds of wet screening, resulting in long cycles, high costs, and strong dependence on experimental conditions.

[0003] With the development of deep learning and structure prediction technologies, AI-assisted protein conjugate design methods have rapidly advanced. For example, some methods can generate candidate conjugate structures de novo based on the target protein structure, while others can perform sequence design or structure optimization based on a given backbone. However, existing deep learning and structure prediction methods still have some shortcomings, including high computational resource consumption, unstable iterative optimization direction, and low automation in the computation process, making it difficult to improve the efficiency of candidate protein conjugate generation. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method and system for iterative optimization of protein conjugates based on intelligent agent collaborative execution, so as to alleviate the above-mentioned technical problems and improve the generation efficiency of candidate protein conjugates.

[0005] In a first aspect, embodiments of the present invention provide a protein conjugate iterative optimization method based on agent-based collaborative execution, applied to an agent execution system, the agent execution system comprising multiple functional agents; the method comprising: acquiring initial input data through a task orchestration agent; wherein the initial input data includes an initial complex structure of the protein conjugate to be optimized and a target protein; determining the complex structure of the current round based on the initial input data; invoking a binding site prediction agent to input the complex structure of the current round into a binding site prediction model, and outputting the binding site probability corresponding to each residue in the protein conjugate to be optimized through the binding site prediction model; invoking a residue locking decision agent to determine whether the residues in the protein conjugate to be optimized are locked, and if so, adding the residues to a locked residue set; and determining, based on the locked residue set and a preset redesign range, a... A set of freely designed sites is defined; wherein the residues in the locked residue set remain unchanged during subsequent sequence design; the set of freely designed sites consists of residues that belong to the redesign scope but not to the locked residue set; a sequence design agent is invoked to design sequences for the residues in the set of freely designed sites using a structure-conditional sequence design model, generating multiple candidate protein-conjugate sequences; a structure evaluation agent is invoked to evaluate the candidate protein-conjugate sequences and the sequence of the target protein, obtaining the corresponding candidate complex structure and the structure confidence index corresponding to the candidate complex structure; a candidate selection agent is invoked to determine the next iteration seed from the multiple candidate protein-conjugate sequences based on the structure confidence index, and the functional agent is scheduled to repeatedly execute the iterative optimization process based on the next iteration seed until a preset stopping condition is met, and then the optimization result is output.

[0006] In conjunction with the first aspect, this embodiment of the invention provides a first possible implementation of the first aspect, wherein the method further includes: after obtaining the initial input data through the task orchestration agent, calculating the initial structure confidence index corresponding to the initial complex structure; determining whether the initial structure confidence index is greater than a preset initial confidence threshold; if so, determining the initial complex structure as the complex structure of the current round; calling the structure preparation agent to read the complex structure of the current round, and submitting the complex structure of the current round to the binding site prediction agent.

[0007] In conjunction with the first aspect, this embodiment of the invention provides a second possible implementation of the first aspect, wherein the step of calling the residue locking decision agent to determine whether the residues in the protein-protein conjugate to be optimized are locked includes: calling the residue locking decision agent to read the binding site probability corresponding to each residue in multiple consecutive iterations; if the binding site probability of the residue in multiple consecutive iterations is greater than or equal to a preset probability threshold, then the residue is determined to be locked, and the residue is added to the locked residue set; or, determining whether the residue is a pre-set permanently locked residue; if so, determining that the residue is locked, and adding the residue to the locked residue set; or, determining whether the residue is outside the redesign range; if so, determining that the residue is locked, and adding the residue to the locked residue set.

[0008] In conjunction with the second possible implementation of the first aspect, this embodiment of the invention provides a third possible implementation of the first aspect, wherein the method further includes: obtaining the residue-level structure confidence of the residue before adding the residue to the locked residue set; if the binding site probability of the residue is greater than or equal to the preset probability threshold, and the residue-level structure confidence of the residue is greater than or equal to the preset structure confidence threshold, then the residue is determined as a candidate locked residue; and the residue is added to the locked residue set based on the candidate locked residues in consecutive iterations.

[0009] In conjunction with the first aspect, this invention provides a fourth possible implementation of the first aspect, wherein the step of using a structural conditional sequence design model to sequence design the residues in the set of freely designed sites to generate multiple candidate protein conjugate sequences includes: inputting the complex structure corresponding to the current iteration, the protein conjugate sequence of the current iteration, the locked residue set, and the set of freely designed sites into the structural conditional sequence design model; and using the structural conditional sequence design model to sample or replace amino acids of the residues in the set of freely designed sites while keeping the amino acids corresponding to the residues in the locked residue set unchanged, thereby generating multiple candidate protein conjugate sequences.

[0010] In conjunction with the first aspect, this embodiment of the invention provides a fifth possible implementation of the first aspect, wherein the step of calling the structure evaluation agent to evaluate the candidate protein conjugate sequence and the target protein sequence includes: calling the structure evaluation agent to input the candidate protein conjugate sequence and the target protein sequence into the evaluation model, and obtaining the corresponding candidate complex structure and the structure confidence index corresponding to the candidate complex structure through the evaluation model.

[0011] In conjunction with the first aspect, this embodiment of the invention provides a sixth possible implementation of the first aspect, wherein the step of invoking the candidate selection agent to determine the next iteration seed from a plurality of candidate protein-protein conjugate sequences based on the structural confidence index includes: invoking the candidate selection agent to sort the plurality of candidate protein-protein conjugate sequences according to the structural confidence index; determining the candidate protein-protein conjugate sequences whose sorting results satisfy preset selection conditions as the next iteration seed; wherein the preset selection conditions include at least one of the following: highest structural confidence index, structural confidence index exceeding a preset candidate threshold, or highest comprehensive score; the comprehensive score is obtained by weighting at least two of a plurality of indicators including the structural confidence index.

[0012] In conjunction with the first aspect, this invention provides a seventh possible implementation of the first aspect, wherein the method further includes: the invocation candidate selection agent feeding back the next round iteration seed to the task orchestration agent, so that the task orchestration agent schedules the functional agent to repeatedly execute the iterative optimization process based on the next round iteration seed, until a preset stopping condition is met, and then outputs the optimization result; wherein the preset stopping condition includes at least one of the following: the optimal structure confidence index of the current round reaches a preset absolute threshold; the gain of the structure confidence index between two adjacent iterations is less than a preset gain threshold; no improvement in the structure confidence index is obtained for a consecutive preset number of rounds; the maximum number of iteration rounds is reached; the number of free design sites in the free design site set is less than a preset number threshold; the number of locked residues in the locked residue set reaches a preset number threshold; the coverage rate of the locked residues reaches a preset coverage rate threshold.

[0013] In conjunction with the first aspect, this invention provides an eighth possible implementation of the first aspect, wherein the protein conjugate to be optimized is a short peptide conjugate; the preset redesign range is all residues of the short peptide conjugate, and no permanently locked residues are set; or, the protein conjugate to be optimized is a nanobody; the preset redesign range is the complementarity-determining region of the nanobody; the framework region residues of the nanobody are located outside the redesign range, and / or the preset anchoring points in the complementarity-determining region are set as permanently locked residues; or, the protein conjugate to be optimized is a microprotein conjugate; the preset redesign range is the entire chain of the microprotein conjugate or a specified preset residue interval; if the entire chain optimization mode is adopted, no permanently locked residues are set; if a specific backbone or functional site is retained, the corresponding residues are set as permanently locked residues.

[0014] Secondly, embodiments of the present invention also provide an intelligent agent execution system, the intelligent agent execution system comprising multiple functional intelligent agents to execute the method described in the first aspect.

[0015] Thirdly, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect, and / or to run the intelligent agent execution system of the second aspect.

[0016] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method described in the first aspect and / or causes a computing device to run the intelligent agent execution system of the second aspect.

[0017] The embodiments of the present invention bring the following beneficial effects: This invention provides a method and system for iterative optimization of protein conjugates based on agent-based collaborative execution. This method and system can be applied to an agent-based execution system, which includes multiple functional agents. Specifically, initial input data can be obtained through a task orchestration agent; the complex structure for the current round can be determined based on the initial input data; a binding site prediction agent is invoked to input the complex structure for the current round into a binding site prediction model, which outputs the binding site probability corresponding to each residue in the protein conjugate to be optimized; a residue locking decision agent is invoked to determine whether the residues in the protein conjugate to be optimized are locked; if so, the residues are added to the locked residue set; and, based on the locking... The process involves defining a set of residues and a predefined redesign range to determine the set of free design sites. A sequence design agent is invoked to design sequences from the residues in the free design site set using a structure-conditional sequence design model, generating multiple candidate protein-conjugate sequences. A structure evaluation agent is invoked to evaluate the candidate protein-conjugate sequences against the target protein sequence, obtaining the corresponding candidate complex structures and their corresponding structure confidence indices. A candidate selection agent is invoked to determine the next iteration seed from the multiple candidate protein-conjugate sequences based on the structure confidence indices. The iterative optimization process is then repeated using an agent that schedules the next iteration seed until a predefined stopping condition is met, at which point the optimization result is output. This entire iterative process, through the collaboration of multiple agents, enables model invocation and iterative optimization, thereby improving the optimization efficiency, process interpretability, and structure confidence of protein conjugates, ultimately reducing computational costs and increasing the generation efficiency of candidate protein conjugates.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 A flowchart of an iterative optimization method for protein conjugates based on agent-cooperative execution is provided for embodiments of the present invention; Figure 2 A general flowchart of multi-agent coordinated execution provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent agent execution system provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Currently, with the development of deep learning and structure prediction technologies, AI-assisted protein conjugate design methods have rapidly advanced. However, existing deep learning-based structure prediction technologies still have the following shortcomings: First, existing de novo generation methods typically use single-batch generation or scaling up the sampling scale as the main optimization approach. For protein conjugates that already have reasonable starting structures, blindly scaling up the generation scale will significantly increase computational resource consumption, and may not necessarily fully tap the optimization potential of the existing starting structures.

[0024] Second, while some existing sequence design methods can generate sequences based on the structural backbone, they often lack guidance on optimization directions at the residue level during iterative use. If all variable residues are designed in an approximately uniform random manner in each iteration, it is easy to disturb key residues that have formed stable binding interactions, resulting in unstable optimization directions, an excessively large search space, and large fluctuations in results.

[0025] Third, existing structure optimization or sequence optimization methods often only achieve one aspect of structure awareness or residue orientation. For example, some methods can use structural information for sequence design but lack a mechanism for judging residue importance round by round; other methods can provide binding-related signals at the residue level but fail to translate these signals into stable sequence design constraints. Therefore, current technology still lacks an optimization method that can simultaneously utilize structural information and residue-level prediction information and gradually stabilize the binding interface during the iteration process.

[0026] Fourth, in the process of protein binding optimization, the prediction results of a single-round model may be affected by structural noise, model uncertainty, or local conformational fluctuations. If residues are fixed directly based on the single-round prediction results, non-critical residues may be incorrectly locked; if no locking is performed at all, the search space cannot be effectively narrowed and critical binding residues cannot be protected.

[0027] Fifth, existing protein conjugate optimization processes typically require the manual cascading of multiple heterogeneous models and tools, such as structure preparation, binding site prediction, sequence generation, complex evaluation, candidate sorting, and failure retries. The lack of a unified task state memory, model call scheduling, and result verification mechanism between different steps results in insufficient automation, weak anomaly recovery capabilities, and difficulty in collaborative exploration across multiple candidate branches.

[0028] Therefore, there is an urgent need for an iterative optimization method for protein conjugates that can achieve efficient refinement of protein conjugates with low computational cost.

[0029] Based on this, the embodiments of the present invention provide a method and system for iterative optimization of protein conjugates based on intelligent agent collaborative execution, which can alleviate the above-mentioned technical problems and improve the generation efficiency of candidate protein conjugates.

[0030] To facilitate understanding of this embodiment, a detailed description of the protein conjugate iterative optimization method based on intelligent agent cooperative execution disclosed in this embodiment of the invention will be provided first.

[0031] In one possible implementation, embodiments of the present invention provide an iterative optimization method for protein conjugates based on agent-based cooperative execution, applied to an agent execution system, the agent execution system comprising multiple functional agents; specifically, as shown... Figure 1The flowchart shown illustrates an iterative optimization method for protein conjugates based on agent-cooperative execution. This method includes the following steps: Step S102: Obtain initial input data through the task orchestration agent; In this embodiment of the invention, the initial input data includes the initial complex structure of the protein conjugate to be optimized and the target protein; Specifically, the protein conjugates in the embodiments of the present invention can be short peptides, polypeptides, nanobodies, antibody fragments, microproteins, de novo designed proteins, or other protein molecules capable of binding to target proteins.

[0032] Furthermore, the initial input data may include at least one of the following: the amino acid sequence of the target protein; the three-dimensional structure of the target protein; the amino acid sequence of the protein-to-be-optimized protein conjugate; the initial complex structure formed by the protein-to-be-optimized protein conjugate and the target protein; the predicted structure of the initial complex formed by the protein-to-be-optimized protein conjugate and the target protein; the initial structure confidence index; the redesign range; the permanently locked residue set; the prohibited amino acid set; the maximum number of iterations; the binding site probability threshold; the structure confidence threshold; stopping condition parameters, etc. All of this data can be written into the shared state storage space. The various functional agents of the intelligent agent execution system can read the data in the shared state storage space, execute the corresponding model call or decision rule, and write the output results back to the shared state storage space.

[0033] Furthermore, the information stored in the shared state storage space in this embodiment of the invention may also include at least one of the following: current round number, current seed sequence, current complex structure, residue-level binding site probability, residue-level structure confidence, locked residue set, freely designed site set, candidate sequence set, candidate complex structure, structure confidence index, branching state, failure reason, number of retries, and stopping condition judgment result.

[0034] Step S104: Determine the complex structure of the current round based on the initial input data; Step S106: Call the binding site prediction agent, input the complex structure of the current round into the binding site prediction model, and output the binding site probability corresponding to each residue in the protein complex to be optimized through the binding site prediction model. Step S108: Invoke the residue locking decision agent to determine whether the residues in the protein-protein conjugate to be optimized are locked. If so, add the residues to the locked residue set. Also, determine the free design site set based on the locked residue set and the preset redesign range. Among them, the residues in the locked residue set remain unchanged during subsequent sequence design; the free design site set consists of residues that belong to the redesign scope and do not belong to the locked residue set; Step S110: Invoke the sequence design agent and use the structural conditional sequence design model to design sequences of residues in the set of free design sites to generate multiple candidate protein-binding sequences; Step S112: Call the structure evaluation agent to evaluate the candidate protein conjugate sequence and the target protein sequence to obtain the corresponding candidate complex structure and the structure confidence index corresponding to the candidate complex structure. Step S114: Invoke the candidate selection agent to determine the next iteration seed from multiple candidate protein conjugate sequences based on the structure confidence index. The iterative optimization process is repeated by scheduling the agent based on the next iteration seed until the preset stopping condition is met, and then the optimization result is output.

[0035] In practical use, to make subsequent iterations more reasonable, in this embodiment of the invention, after obtaining the initial input data through the task orchestration agent and before determining the complex structure of the current round, the initial structure confidence index corresponding to the initial complex structure can be calculated; it can be determined whether the initial structure confidence index is greater than the preset initial confidence threshold; if so, the initial complex structure is determined as the complex structure of the current round; the structure preparation agent is called to read the complex structure of the current round, and the complex structure of the current round is submitted to the binding site prediction agent.

[0036] In this embodiment of the invention, the initial structure confidence index is used to determine whether the initial complex structure or predicted structure meets the initial optimization conditions. If the initial structure confidence index meets a preset initial confidence threshold, the initial complex structure or predicted structure is used as the starting structure for subsequent iterative optimization.

[0037] Furthermore, in this embodiment of the invention, the initial structure confidence index includes at least one of the following: interface prediction template modeling score, prediction template modeling score, prediction local distance difference test score, number of interface contacts, root mean square deviation of the interface, binding energy score, interface area, number of hydrogen bonds, and number of hydrophobic contacts. Specifically, at least one of these can be selected as the initial structure confidence index, depending on the actual application; this embodiment of the invention does not impose any limitations on this selection.

[0038] For example, the initial structure confidence index can be the interface prediction template modeling score. If the initial structure confidence index is greater than a preset initial confidence threshold, such as greater than 0.7, then the initial complex structure is considered a reasonable starting structure, and subsequent iterative refinement is performed on it. It should be noted that the initial confidence threshold can be adjusted according to different application scenarios and does not constitute a limitation on the scope of protection of this invention.

[0039] Furthermore, once the structure of the complex in the current round is determined, subsequent functional agents can be invoked, such as binding site prediction agents, residue locking decision agents, sequence design agents, structure evaluation agents, and candidate selection agents, to perform collaborative work among multiple functional agents, thereby executing the protein conjugate iterative optimization method based on agent collaborative execution in this embodiment of the invention.

[0040] For ease of understanding, Figure 2 A general flowchart of multi-agent coordinated execution is shown, and, Figure 2 The diagram shows a task orchestration agent 20, a structure preparation agent 21, a binding site prediction agent 22, a residue locking decision agent 23, a sequence design agent 24, a structure evaluation agent 25, and a candidate selection agent 26, respectively. Figure 2 The document also shows a shared state storage space 27. It should be understood that... Figure 2 The intelligent agents shown are only some of the intelligent agents in the intelligent agent execution system of this embodiment of the invention. In other embodiments, other intelligent agents may be included, such as output intelligent agents, etc. The specific implementation shall be subject to the actual use, and this embodiment of the invention does not impose any restrictions on this.

[0041] Furthermore, for ease of understanding, the following will be combined with... Figure 2 The overall flowchart of multi-agent coordinated execution shown further illustrates the protein conjugate iterative optimization method based on agent cooperative execution provided in this embodiment of the invention, including the following processes: (1) Obtain initial input data through a task orchestration agent; This process corresponds to step S102 above. In this process, after the task orchestration agent obtains the initial input data, it establishes a shared state storage space for recording the current iteration state. The shared state storage space includes at least the current seed sequence, the current complex structure, the locked residue set, the set of freely designed sites, the candidate sequence set, and the structure confidence index, etc.

[0042] After further determining that the initial complex structure is a reasonable starting structure based on the initial structure confidence index, subsequent iterative refinement is performed.

[0043] (2) The structure preparation agent determines the complex structure of the current round based on the initial complex structure or the predicted structure, and submits the complex structure of the current round to the binding site prediction agent; (3) Predict the probability of binding sites at the residue level by predicting the binding site of the agent; Specifically, this process corresponds to step S106 above. In practice, the complex structure of the protein conjugate and target protein corresponding to the current iteration can be input into the binding site prediction model. The binding site prediction model outputs the binding site probability corresponding to each residue in the protein conjugate to be optimized. In this embodiment, the binding site probability is used to characterize the likelihood that the corresponding residue will participate in the binding interface between the protein conjugate to be optimized and the target protein. The binding site prediction model can output the binding site probability at the residue level based on the protein structure and is insensitive to the identity of the chaperone protein at the binding interface.

[0044] In one feasible implementation, the aforementioned binding site prediction model can be a machine learning model or a deep learning model capable of outputting the probability of binding sites at the residue level based on protein structure. For example, the ScanNet model can be used to process the current complex structure to obtain the binding site probability of each residue in the complex chain. It should be understood that ScanNet is only one optional implementation, and any model capable of outputting residue-level binding-related probabilities, scores, or priorities can be used as a binding site prediction model; this embodiment of the invention does not impose any limitations on this.

[0045] For example, in the t-th iteration, suppose the protein conjugate to be optimized contains N residues, and the probability of the binding site corresponding to the i-th residue is denoted as p_i(t), where i is the residue number and t is the iteration number. When p_i(t) is greater than or equal to the preset probability threshold tau, the i-th residue is marked as a candidate binding site residue in the t-th iteration.

[0046] (4) Residue locking is performed by a residue locking decision-making agent based on multiple rounds of consensus; Specifically, this process corresponds to the above. Figure 1 Step S108. In this process, it is achieved by invoking a residue-locking decision agent. Furthermore, in this embodiment of the invention, locked residues from different sources are determined, typically including the following three categories: The first type is dynamically locked residues, which are determined by consensus on binding site probabilities across multiple iterations. In this embodiment, these locked residues are automatically generated during the iteration process. Dynamically locked residues reflect residues that continuously and stably participate in the binding interface under the current structural environment. Specifically, these dynamically locked residues are implemented by a residue locking decision agent, including: invoking the residue locking decision agent, which can read the binding site probabilities corresponding to each residue in multiple iterations from the shared state storage space; and determining locked residues based on the binding site probabilities in multiple iterations. Specifically, if the binding site probability of a residue in multiple iterations is greater than or equal to a preset probability threshold, then the residue is determined to be locked, and the residue is added to the locked residue set.

[0047] For example, a two-round consensus locking rule can be preferred. Specifically, for the i-th residue, it is determined whether the binding site probability of the residue in the (t-1)-th iteration is greater than or equal to a preset probability threshold, and whether the binding site probability of the residue in the t-th iteration is greater than or equal to the preset probability threshold. If the binding site probability of the residue in the (t-1)-th iteration and the binding site probability in the t-th iteration are both greater than or equal to the preset probability threshold, then the residue is added to the locked residue set. That is, if the binding site probability of the residue satisfies p_i(t-1)≥tau and p_i(t)≥tau, then the i-th residue is added to the locked residue set. This two-round consensus locking rule can also be called the C-2 locking rule.

[0048] Where t is an integer greater than or equal to 1; the above two consecutive rounds of consensus locking rules are used to treat the single-round prediction results as evidence rather than immediately locking commitments, thereby filtering out the influence of single-round prediction noise and local conformational fluctuations. That is, only when the same residue is predicted to have a binding site probability higher than the preset probability threshold in two consecutive rounds is it considered to be a relatively stable binding-related residue and kept fixed in subsequent iterations.

[0049] The second category is permanently locked residues, and the third category is out-of-scope locked residues. Permanently locked residues are pre-specified by the user or configuration files. For example, in nanobody optimization scenarios, framework region residues, structural anchor sites, conserved residues, cysteine ​​residues, or other residues that play a crucial role in backbone stability, disulfide bond constraint, or functional maintenance can be set as permanently locked residues. Permanently locked residues remain fixed throughout the iteration process. Out-of-scope locked residues refer to residues located outside the pre-defined redesign range. These residues, even if not set as permanently locked residues, do not participate in sequence design and are therefore considered locked residues in the algorithm implementation.

[0050] Therefore, in this embodiment of the invention, when determining whether a residue is locked, it may be further determined whether the residue is a pre-set permanently locked residue; if so, the residue is determined to be locked and the residue is added to the locked residue set; or, it may be determined whether the residue is outside the redesign range; if so, the residue is determined to be locked and the residue is added to the locked residue set.

[0051] In practice, the set of candidate residues that meet the preset probability threshold in round t can be denoted as C_t. Then, the set of dynamically locked residues added in round t, D_t, is the intersection of the candidate residue set C_t in round t and the candidate residue set C_(t-1) in round t-1. The locked set L_t in round t is obtained by merging the locked set L_(t-1) in the previous round, the dynamically locked residue set D_t, the permanently locked residue set P, and the out-of-range locked residue set O. Here, L_t represents the locked set in round t, L_(t-1) represents the locked set in the previous round, D_t represents the dynamically locked residue set obtained based on consensus from two consecutive rounds, P represents the preset set of permanently locked residues, and O represents the set of out-of-range locked residues.

[0052] By using the above three types of residue locking sources, this embodiment of the invention can automatically identify binding hotspots that gradually stabilize during the iteration process, retain key structural constraints specified by the user based on prior knowledge, and limit the optimization region according to task requirements.

[0053] Furthermore, the locked residue set in this embodiment of the invention is a monotonically increasing set; wherein, residues that have been added to the locked residue set remain locked in subsequent iterations and are no longer released as free design sites; this monotonically increasing mechanism prevents residues that have been stably involved in the binding interface from being repeatedly perturbed in subsequent sequence design.

[0054] Furthermore, in this embodiment of the invention, before adding a residue to the locked residue set, the residue-level structural confidence level of the residue can be obtained; structural confidence gating is then performed on the residue based on the residue-level structural confidence level. Specifically, if the binding site probability of a residue is greater than or equal to a preset probability threshold, and the residue-level structural confidence level of the residue is greater than or equal to a preset structural confidence threshold, then the residue is determined as a candidate locked residue; the residue is added to the locked residue set based on the candidate locked residues in multiple consecutive iterations.

[0055] Specifically, a residue is allowed to enter the candidate locking set only when it simultaneously satisfies both the preset probability threshold for binding site probability and the structural confidence threshold. For example, for the i-th residue, if it satisfies p_i(t)>=tau and q_i(t)>=tau_q in round t, it is marked as a candidate locking residue in round t. Here, q_i(t) represents the residue-level structural confidence of the i-th residue in round t, and tau_q represents the preset probability threshold at this time, i.e., the preset structural confidence threshold. When the residue is identified as a candidate locking residue in multiple consecutive iterations, it is determined that the residue is added to the locking residue set. Furthermore, the residue-level structural confidence in this embodiment can be a predicted local distance difference test score or other residue-level structural reliability indicators. Moreover, the structural confidence gating process corresponding to the residue-level structural confidence is used to prevent false positive binding site prediction results in low-confidence regions from being incorrectly locked, thereby further improving the stability of the locking decision.

[0056] (5) Determine the free design site; This process corresponds to the above. Figure 1 In step S108, specifically, after determining the locked residue set for the current round, the unlocked residues within the preset redesign range are identified as free design sites, and these residues are added to the free design site set. The free design site set shrinks as the locked residue set increases, causing the sequence search space to gradually decrease during iteration, thereby reducing computational resource consumption.

[0057] For example, let A be the complete set of residues in a protein-protein complex, R be the set of residues within a pre-defined redesign range, and L_t be the set of locked residues in round t. Then, the set of free design sites F_t in round t can be expressed as F_t = R - L_t; where F_t represents the set of free design sites in round t, R represents the set of residues within the pre-defined redesign range, and R - L_t represents the set of residues obtained by removing the locked set L_t from the pre-defined redesign range R. In other words, only residues within the redesign range that are not locked are allowed to change in the current round of sequence design. As iterations proceed, the set of locked residues gradually expands, while the set of free design sites gradually shrinks, thereby reducing the search space for subsequent sequence designs.

[0058] (6) Sequence design agents design sequences based on freely designed sites; This process corresponds to the above. Figure 1 The process of step S110 in the process.

[0059] Specifically, the complex structure corresponding to the current iteration, the protein conjugate sequence of the current iteration, the locked residue set, and the free design site set can be input into the structural conditional sequence design model. The structural conditional sequence design model can sample or replace amino acids of residues in the free design site set while keeping the amino acids corresponding to residues in the locked residue set unchanged, thereby generating multiple candidate protein conjugate sequences.

[0060] In practice, the sequence design agent reads the complex structure of the current iteration, the protein conjugate sequence of the current iteration, the locked residue information, and the free design site information from the shared state storage space, and inputs the above information into the structure-conditional sequence design model to design sequences for the free design sites and generate multiple candidate protein conjugate sequences.

[0061] In a preferred embodiment, a pre-trained structure-conditional sequence design model, such as the ProteinMPNN model, is used. The ProteinMPNN model can generate candidate amino acids for freely designed sites based on a given protein backbone structure and fixed residue constraints. It should be understood that the ProteinMPNN model is only one possible implementation; other models capable of protein sequence design under structural constraints can also be used in this embodiment of the invention, and this embodiment does not impose any limitations on them.

[0062] Furthermore, in this embodiment of the invention, the above sequence design process can set a prohibited amino acid set. Specifically, a pre-set prohibited amino acid set can be obtained, and the optional amino acid types of some free design sites in the free design site set can be restricted based on this prohibited amino acid set. For example, in a specific nanobody complementarity determination region optimization task, certain amino acids that are prone to introducing instability or undesirable chemical properties can be prohibited from appearing at the free design site. Moreover, each iteration can generate a preset number of candidate sequences. For example, 500, 1000, 2000, or other numbers of candidate protein-protein conjugate sequences can be generated for each iteration branch. The number of candidate sequences can be adjusted according to the protein-protein conjugate length, redesign range size, computational resources, and task difficulty; this embodiment of the invention does not impose any limitations on this.

[0063] (7) Perform structural prediction or structural evaluation through a structural evaluation agent; Specifically, this process corresponds to step S112 above. Specifically, the structure evaluation agent can be invoked to input the candidate protein-conjugate sequence and the target protein sequence into the evaluation model. The evaluation model then obtains the corresponding candidate complex structure and its corresponding structure confidence index. Furthermore, the candidate complex structure and its corresponding structure confidence index obtained during this process can be written into the shared state storage space.

[0064] In specific implementation, the evaluation model for this process in the embodiments of the present invention can be a structure prediction model or a structure evaluation model, preferably the Boltz-2 model. The Boltz-2 model can be used to predict or evaluate the structure of the complex between the candidate protein conjugate and the target protein, and output an interface-related confidence index, i.e., the structure confidence index in the embodiments of the present invention. It should be understood that the Boltz-2 model is only one optional implementation of the embodiments of the present invention, and other models capable of predicting or evaluating the structure of protein complexes can also be used in the embodiments of the present invention, and the embodiments of the present invention do not limit this.

[0065] Furthermore, the structural confidence index in the embodiments of the present invention may include, but is not limited to: interface prediction template modeling score, prediction template modeling score, prediction local distance difference test score, number of interface contacts, root mean square deviation of interface, binding energy score, interface area, number of hydrogen bonds, number of hydrophobic contacts, or a combination thereof.

[0066] Furthermore, the structural confidence index in this embodiment of the invention can be used to determine whether the initial complex structure or predicted structure meets the initial optimization conditions, and it can also be used to rank and screen candidate protein-conjugate sequences. For different types of protein-conjugate tasks, different structural confidence indices can be selected as the main ranking criteria, or multiple structural confidence indices can be combined to obtain a comprehensive evaluation result. For example, in one possible implementation, the interface prediction template modeling score is used as the main ranking index, and the sequence with the highest score is selected as the optimal candidate for the current round from the candidate protein-conjugate sequences based on this index. The specific settings can be made according to the actual usage, and this embodiment of the invention does not impose any limitations on this.

[0067] (8) Update the iterative seed by selecting candidate agents; This process corresponds to the above. Figure 1In step S114, specifically, the process involves invoking a candidate selection agent to sort multiple candidate protein-protein conjugate sequences based on a structural confidence index; the candidate protein-protein conjugate sequences whose sorting results meet preset selection criteria are then identified as seeds for the next iteration; wherein the preset selection criteria include at least one of the following: highest structural confidence index, structural confidence index exceeding a preset candidate threshold, or highest comprehensive score; the comprehensive score is obtained by weighting at least two of the multiple indices including the structural confidence index.

[0068] For example, the candidate protein-conjugate sequence with the highest structural confidence index in the current round can be selected as the seed for the next round of iteration; alternatively, considering factors such as structural confidence index, similarity between the candidate protein-conjugate sequence and the previous round's candidate protein-conjugate sequence, structural stability of the candidate complex, diversity of the candidate protein-conjugate sequence, and contact between the candidate structure and the target interface, one or more candidate protein-conjugate sequences can be selected as the seed for the next round of iteration; yet another example is that the seed for the next round of iteration can be selected based on a comprehensive score. This comprehensive score can be obtained by weighting one or more of the following: structural confidence index, interface contact index, sequence diversity index, structural stability index, and similarity index with the previous round's sequence. The weights of different indicators can be configured according to the protein-conjugate type, target protein type, computational resources, or optimization objectives. For example, the interface prediction template modeling score can be used as the main scoring item, and the predicted local distance difference test score, the number of interface contacts, the binding energy score, and the sequence diversity index can be used as auxiliary scoring items to obtain a comprehensive score for candidate ranking. The specific comprehensive score can be set according to actual usage, and this embodiment of the invention does not impose any limitations on this.

[0069] Furthermore, in this embodiment of the invention, a continuously advancing update strategy can also be adopted. That is, even if the optimal candidate protein binding sequence in the current round does not achieve a preset improvement relative to the initial seed, the optimal candidate protein binding sequence in the current round can still be used as the seed for the next round of iteration. This update strategy treats each round as an iterative step along the current optimization direction, rather than requiring each round to exceed the global initial structure.

[0070] (9) Explore the agent's execution bundle search strategy through branching; That is, in the embodiments of the present invention, except Figure 2 The various functional agents shown also include branch exploration agents. Specifically, in this embodiment of the invention, a bundle search strategy executed collaboratively by multiple branch exploration agents is used to maintain multiple candidate iterative branches.

[0071] For example, let the bundle width be B, i.e., then B candidate iteration branches are maintained simultaneously. Here, B is an integer greater than or equal to 2. Each candidate iteration branch can be maintained by a branch exploration agent or a branch task instance, and has its own current seed sequence, current complex structure, current lock set, and current structure confidence index.

[0072] For each candidate iteration branch, in each iteration, the task orchestration agent schedules the corresponding branch exploration agent to independently execute procedures such as predicting the probability of binding sites, determining the set of locked residues, determining the set of freely designed sites, sequence design, and structure evaluation. Based on the structure confidence index corresponding to each candidate iteration branch, the next iteration seed of the candidate iteration branch is updated, and the state of each branch is written into the shared state storage space.

[0073] For example, in one feasible implementation, the aforementioned bundle width B can be set to 3. That is, by performing iterative optimization in parallel or serially through three candidate iterative branches, a certain degree of sequence diversity can be preserved while keeping computational costs under control.

[0074] Furthermore, in this embodiment of the invention, the above-mentioned bundle search strategy also includes a branch dehomogenization step executed by the branch coordination agent. Specifically, to prevent different candidate iteration branches from converging to similar sequences too early, the branch coordination agent can calculate the sequence difference between the protein-binding sequences corresponding to any two candidate iteration branches. This sequence difference can be determined by at least one of Hamming distance, edit distance, or sequence similarity, or by other indicators that can characterize the degree of difference between the two protein-binding sequences. If the sequence difference is less than a preset difference threshold, the branch coordination agent sends a perturbation command to the candidate iteration branch with a lower structural confidence index to perturb at least one free design site in the candidate iteration branch. This perturbation may include randomly replacing amino acids at at least one free design site in the branch, or resampling some free design sites in the branch, in order to increase the sequence diversity between different candidate iteration branches and reduce the risk of getting trapped in local optima.

[0075] (10) Determine the stopping condition through the task orchestration agent; Furthermore, in this process, the candidate selection agent is invoked to feed back the next round of iteration seeds to the task orchestration agent, so that the task orchestration agent can repeatedly execute the iterative optimization process based on the next round of iteration seeds to the scheduling agent until the preset stopping condition is met, and then output the optimization results; that is, the binding site prediction agent, residue locking decision agent, sequence design agent, structure evaluation agent and candidate selection agent are rescheduled to repeatedly execute the processes of predicting the binding site probability, determining the locked residue set, determining the free design site set, sequence design and structure evaluation, etc., until the preset stopping condition is met.

[0076] The preset stop conditions in the embodiments of the present invention include at least one of the following: The confidence index of the optimal structure in the current round reaches the preset absolute threshold; for example, when the interface prediction template modeling score of the candidate complex reaches or exceeds the preset value, it is considered that the current candidate complex structure has reached the confidence of the target structure and the iteration can be stopped. The gain of the structure confidence index between two adjacent iterations is less than a preset gain threshold; for example, when the difference between the confidence of the current optimal structure and the confidence of the previous optimal structure is less than a preset threshold, it is considered that the benefit of continuing the iteration is low and the iteration can be stopped. If the structure confidence index does not improve within a preset number of consecutive rounds; for example, if the best structure confidence index does not exceed the historical best value or the improvement is less than the preset threshold in several consecutive rounds, the patience stop condition is triggered. The maximum number of iteration rounds can be reached; for example, the maximum number of iteration rounds can be set to 5 rounds, 10 rounds, 20 rounds or other values, depending on the actual use case, and the embodiments of the present invention do not impose any restrictions on this.

[0077] If the number of freely designed sites in the set of freely designed sites is less than a preset threshold, or if the number of locked residues in the set of locked residues reaches a preset threshold, it indicates that most residues have been locked and there is insufficient remaining design space, and the iteration can be stopped. The coverage of locked residues reaches the preset coverage threshold; when the coverage ratio of locked residues reaches the coverage threshold, it indicates that the key sites of the binding interface have been basically stabilized and the iteration can be stopped.

[0078] In practice, if the above stopping condition is not met, the process returns to (3) above and enters the next iteration; if the stopping condition is met, the process proceeds to (11) below.

[0079] (11) Output optimization results through the output agent; In practice, the output agent can output the optimized protein conjugate sequence, the optimized protein conjugate structure, or the optimized protein conjugate-target protein complex structure.

[0080] In one possible implementation, the following information may also be output: the structural confidence index for each iteration; the newly added locked residue set for each iteration; the final locked residue set; the final set of freely designed sites; the optimal sequence for each candidate iteration branch; the sequence differences between candidate iteration branches; the sequence change sites before and after optimization; the comparison of the complex structure before and after optimization; and the changes in interface contact before and after optimization.

[0081] Through the above output, users can not only obtain the final optimization results, but also track the process of residue locking, sequence changes and structural confidence changes in each iteration, thereby improving the interpretability of the method.

[0082] Furthermore, in this embodiment of the invention, the protein conjugate to be optimized is preferably a short peptide conjugate; the preset redesign range is all residues of the short peptide conjugate, and no permanently locked residues are set; or, the protein conjugate to be optimized is preferably a nanobody; the preset redesign range is the complementarity-determining region of the nanobody; the framework region residues of the nanobody are located outside the redesign range, and / or the preset anchoring points in the complementarity-determining region are set as permanently locked residues; or, the protein conjugate to be optimized is preferably a micro protein conjugate; the preset redesign range is the entire chain of the micro protein conjugate or a specified preset residue interval; if the entire chain optimization mode is adopted, no permanently locked residues are set; if a specific backbone or functional site is retained, the corresponding residues are set as permanently locked residues.

[0083] Specifically, for ease of understanding, the following will further illustrate this using examples of short peptide conjugates, nanobodies, and micro protein conjugates as the protein conjugates to be optimized: (a) Short peptide conjugate refinement mode: In this embodiment of the invention, the short peptide conjugate can be a polypeptide designed for the target protein or a natural product-derived polypeptide.

[0084] Specifically, the protein conjugate to be optimized is a short peptide conjugate. In this case, the preset redesign scope covers all residues of the short peptide conjugate, and no permanently locked residues are set. This setting allows optimization across the entire chain of the short peptide conjugate, but as iterations proceed, residues predicted as stable binding sites in two consecutive rounds will be gradually locked, thereby protecting residues that have already formed stable binding interactions.

[0085] The specific process is as follows: First, obtain the initial complex structure of the short peptide conjugate and the target protein, and use a structure evaluation model to obtain the initial structure confidence index; second, use a binding site prediction model to predict the binding site probability of each residue in the short peptide chain, and obtain the binding site probability of each residue participating in the binding interface; then, according to the C-2 locking rule, add residues that meet the preset probability threshold in two consecutive rounds to a dynamic locked residue set; next, use the unlocked residues as free design sites, and use a structure conditional sequence design model to generate multiple candidate short peptide sequences; subsequently, evaluate the structure of the complex formed by the candidate short peptide sequence and the target protein, and select the seed for the next iteration based on the structure confidence index; finally, repeat the above process until the stopping condition is met, and output the optimized short peptide sequence or short peptide-target protein complex structure.

[0086] Furthermore, the embodiments of the present invention are applicable to the refinement of short peptide conjugates with existing reasonable starting structures. For example, the target protein may include immune checkpoint proteins, tumor-associated antigens, receptor proteins, enzymes, or other target proteins with drug development value, subject to actual use, and the embodiments of the present invention do not impose any limitations on this.

[0087] Furthermore, in the short peptide conjugate refinement task of this invention, short peptide conjugates corresponding to multiple target proteins are used as initial seeds. After multiple rounds of iterative optimization, the structural confidence index of multiple samples is improved. Experimental results show that there is a negative correlation between the optimization improvement and the initial structural confidence index, that is, samples with low initial structural confidence but still within a reasonable range usually have greater room for improvement.

[0088] (II) Maturation mode of nanobody complementarity-determining region: Specifically, in this embodiment of the invention, the protein conjugate to be optimized is a nanobody; wherein, the nanobody typically includes a framework region and a complementarity-determining region, wherein the complementarity-determining region directly participates in antigen recognition and binding, and the framework region mainly maintains the overall stability of the nanobody backbone. In this embodiment of the invention, the preset redesign range is the complementarity-determining region of the nanobody; the framework region residues of the nanobody are located outside the preset redesign range, and / or the preset anchoring sites in the complementarity-determining region are set as permanently locked residues.

[0089] The complementarity-determining region is preferably the third complementarity-determining region; at least one cysteine ​​residue, structural anchoring residue, conserved residue, or functionally relevant residue pre-specified by the user in the third complementarity-determining region is set as a permanently locked residue to maintain the local conformation of the nanobody, disulfide bond constraints, or key structural features related to antigen recognition; prohibited amino acid sets are also set during sequence design to reduce the possibility of introducing unexpected disulfide bonds, oxidation risks, or hydrophobic aggregation risks in the freely designed sites.

[0090] The specific process is as follows: First, obtain the initial complex structure of the nanobody and the target antigen, and determine the range of the complementarity-determining region (CDR) in the nanobody; second, set the framework region residues as out-of-range locked residues according to the preset configuration, and set the anchor sites, conserved residues, or cysteine ​​residues in the CDR as permanently locked residues; then, use the binding site prediction model to predict the current nanobody-target antigen complex structure at the residue level to obtain the binding site probability of each residue in the CDR; next, use the C-2 locking rule to dynamically lock the residues in the CDR that meet the preset probability threshold for two consecutive rounds; subsequently, use the unlocked residues in the CDR as free design sites, and use the structural conditional sequence design model to generate candidate nanobody sequences.

[0091] In one possible implementation, a set of prohibited amino acids can be set, such as restricting cysteine, methionine, tryptophan, or other amino acids from appearing in the freely designed sites, to reduce the risk of unexpected disulfide bond formation, oxidation, or hydrophobic aggregation. Finally, the structure of the complex formed by the candidate nanobody and the target antigen is evaluated, and the seed for the next iteration is updated according to the structure confidence index until the stopping condition is met, and the optimized nanobody sequence or nanobody-target antigen complex structure is output.

[0092] Through the above methods, the embodiments of the present invention can maintain the stability of the nanobody framework region while performing targeted maturation optimization of the complementarity-determining region, thereby improving the confidence of the binding structure between the nanobody and the target antigen.

[0093] (III) Optimization mode of micro protein conjugates: The miniature protein conjugate can be a protein conjugate generated by de novo protein design methods or a candidate conjugate obtained through other computational design methods. The preset redesign range can be all residues of the miniature protein conjugate or a local residue range specified by the user. If the whole-chain optimization mode is used, no permanently locked residues are set; if it is necessary to retain specific backbones or functional sites, the corresponding residues can be set as permanently locked residues.

[0094] The specific process is as follows: First, obtain the initial complex structure of the microprotein conjugate and the target protein. The initial complex structure can be obtained by de novo design methods, structure prediction models, or docking methods. Second, evaluate the structure of the initial complex structure to obtain an initial structure confidence index. If the initial structure confidence index meets the preset starting conditions, subsequent iterative refinement is performed. Then, use a binding site prediction model to predict the binding site probability of each residue in the microprotein conjugate, and lock the relevant residues stably according to the C-2 locking rule round by round. Next, use a structural conditional sequence design model to design sequences for the unlocked free sites, generating multiple candidate microprotein conjugate sequences. Subsequently, perform structure prediction or evaluation on the complex formed by the candidate microprotein conjugate sequences and the target protein to obtain the candidate structure confidence index, and select the preferred candidate protein conjugate sequence as the seed for the next round of iterations. Finally, after meeting the preset stopping conditions, output the optimized microprotein conjugate sequence or structure.

[0095] Furthermore, in the microprotein conjugate optimization task, this embodiment of the invention uses microprotein conjugates generated by de novo design methods as the starting structure and performs full-chain refinement using the method of this embodiment. During the iteration process, some residues are locked in each round, the remaining free sites are gradually reduced, and finally, candidate microprotein conjugates with higher structural confidence indices are obtained.

[0096] In the maturation task of the complementarity-determining region (CDR) of nanobodies, the framework region of the nanobodies is fixed, and only the CDR is optimized. Some anchor sites, conserved residues, or cysteine ​​residues are set as permanently locked residues. Experimental results show that even under conditions of high initial structure confidence, narrow redesign window, and amino acid usage restrictions, the method of this invention can still achieve improved structure confidence.

[0097] Furthermore, the embodiments of the present invention can serve as a post-optimization module for de novo protein conjugate design, further refining candidate micro-protein conjugates obtained from existing generative models, thereby improving the structural confidence of candidate conjugates without significantly expanding the scale of de novo generative sampling.

[0098] Furthermore, based on the above embodiments, this invention also provides an intelligent agent execution system. Specifically, this intelligent agent execution system can be deployed in a server, workstation, cloud computing platform, or other electronic device. Moreover, the intelligent agent execution system includes multiple functional intelligent agents to execute the above-described... Figure 1 The method shown.

[0099] In practical implementation, multiple functional agents can include task orchestration agents, structure preparation agents, binding site prediction agents, residue locking decision agents, sequence design agents, structure evaluation agents, candidate selection agents, branch exploration agents, and output agents, etc. Furthermore, Figure 3 A schematic diagram of an intelligent agent execution system is shown. In this embodiment, the intelligent agent execution system includes an input and configuration layer 30, a task orchestration layer 31, an agent layer 32, a model and tool layer 33, and an output layer 34. The input and configuration layer can input or configure files, including target proteins, initial or predicted complex structures, redesign ranges, various thresholds, and stopping conditions. The task orchestration layer includes a task orchestration agent, while other functional agents are located in the agent layer. The model and tool layer stores various computational tools and models used. The output layer outputs various information about the optimization process.

[0100] Specifically, the main functions or execution processes of each functional intelligent agent are as follows: The aforementioned task orchestration agent can be used to obtain the initial or predicted structure of the protein-to-target protein complex, read the configuration file, establish a shared state storage space, and schedule other functional agents according to a preset iterative process. The configuration file in this embodiment may include preset redesign ranges, permanently locked residue sets, prohibited amino acid sets, iteration rounds, probability thresholds, structure confidence thresholds, and stopping condition parameters, etc.

[0101] The structure preparation agent is used to generate or organize the complex structure for the current round based on the current seed sequence and target protein information.

[0102] The binding site prediction agent is used to process the complex structure of the current round using the binding site prediction model to obtain the binding site probability of each residue in the protein complex to be optimized.

[0103] The residue-locking decision agent is used to determine locked residues based on the binding site probabilities across multiple consecutive rounds. Specifically, the residue-locking decision agent can determine whether the same residue satisfies a preset probability threshold (a preset threshold for binding site probabilities) in two consecutive rounds. If it does, it is added to a dynamic set of locked residues. The residue-locking decision agent can also merge the dynamic set of locked residues, the permanently locked set of residues, and the out-of-range locked set of residues to obtain the set of locked residues for the current round.

[0104] The free site determination function can be performed by a residue-locking decision agent to determine the free design site based on the locked residue set of the current round and a preset redesign range.

[0105] Sequence design agents are used to design sequences for freely designed sites using structure-conditional sequence design models, generating multiple candidate protein-protein conjugate sequences.

[0106] The structure evaluation agent is used to evaluate candidate complexes formed by candidate protein-binding sequences and target proteins using structure prediction or structure evaluation models, and obtain the confidence index of candidate structures.

[0107] The candidate selection agent is used to select the seed for the next iteration based on the candidate structure confidence index; The task orchestration agent is used to determine whether the preset stopping condition is met. If the stopping condition is not met, the task orchestration agent continues to schedule the binding site prediction agent, residue locking decision agent, sequence design agent, structure evaluation agent, and candidate selection agent to execute the next round of iteration; if the stopping condition is met, the output agent is called.

[0108] The output agent is used to output the optimized protein-protein complex sequence, structure, complex structure, structure confidence index, locked residue set, free design site set, and iteration trajectory information.

[0109] The intelligent agent execution system provided in this embodiment of the invention is based on the same inventive concept as the above method embodiment, and its technical effects are the same as those of the above method embodiment, so they will not be repeated here.

[0110] Furthermore, embodiments of the present invention also provide an electronic device, specifically, such as... Figure 4The diagram illustrates the structure of an electronic device, which includes a memory 40, a processor 41, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the steps of the method described above and / or runs the intelligent agent execution system described above. Figure 4 The image also shows a graphics processor 42.

[0111] In this embodiment of the invention, the electronic device may be a server, workstation, personal computer, cloud computing node, graphics processor server, edge computing device, or other device with data processing capabilities.

[0112] The aforementioned memory may include high-speed random access memory or non-volatile memory, such as disk storage, flash memory devices, or other non-volatile solid-state memory. The processor may be a central processing unit, graphics processing unit, tensor processor, neural network processor, field-programmable gate array, or other processing unit capable of executing computer programs.

[0113] also, Figure 4 The diagram also shows a storage medium 43 connected to the processor, a network interface 44 connected to the graphics processor, and an input / output interface 45 connected to the memory.

[0114] further, Figure 4 The diagram also shows an execution module 46. In one possible implementation, a computer program is stored in the execution module. The computer program includes program instructions for performing the following operations: obtaining an initial complex structure; performing prediction of binding site probabilities; performing residue locking based on two consecutive rounds of consensus; determining freely designed sites; generating candidate binding sequences; evaluating candidate structures; updating the iteration seed; determining stopping conditions; and outputting optimization results, etc.

[0115] Furthermore, embodiments of the present invention also provide a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the protein conjugate iterative optimization method based on agent-based cooperative execution described in any of the above embodiments.

[0116] The computer-readable storage medium may include, but is not limited to: read-only memory, random access memory, disk, optical disk, flash memory, mobile storage device, cloud storage medium or other media capable of storing computer programs, and the embodiments of the present invention do not impose any limitations on this.

[0117] The intelligent agent execution system provided in this embodiment of the invention has the same technical features as the protein conjugate iterative optimization method based on intelligent agent cooperative execution provided in the above embodiment, so it can also solve the same technical problems and achieve the same technical effects.

[0118] The computer program product of the protein conjugate iterative optimization method and system based on intelligent agent cooperative execution provided by the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0119] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0120] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0121] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0122] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0123] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A protein conjugate iterative optimization method based on agent-based collaborative execution, characterized in that, Applied to an intelligent agent execution system, the intelligent agent execution system comprising multiple functional intelligent agents; The method includes: Initial input data is obtained through a task orchestration agent; wherein, the initial input data includes the initial complex structure of the protein conjugate to be optimized and the target protein; The complex structure of the current round is determined based on the initial input data; The binding site prediction agent is invoked, and the complex structure of the current round is input into the binding site prediction model. The binding site prediction model outputs the probability of the binding site corresponding to each residue in the protein complex to be optimized. The residue locking decision agent is invoked to determine whether the residues in the protein-protein conjugate to be optimized are locked. If so, the residues are added to the locked residue set. Based on the locked residue set and a preset redesign range, a free design site set is determined. The residues in the locked residue set remain unchanged during subsequent sequence design. The free design site set consists of residues that belong to the redesign range but not to the locked residue set. The sequence design agent is invoked to design sequences of the residues in the set of free design sites using a structural conditional sequence design model, generating multiple candidate protein-conjugate sequences. The structure evaluation agent is invoked to evaluate the sequence of the candidate protein conjugate and the sequence of the target protein, thereby obtaining the corresponding candidate complex structure and the structure confidence index corresponding to the candidate complex structure. The candidate selection agent is invoked to determine the next iteration seed from multiple candidate protein-protein conjugate sequences based on the structure confidence index. The functional agent is then scheduled to repeatedly execute the iterative optimization process based on the next iteration seed until a preset stopping condition is met, and the optimization result is then output.

2. The method according to claim 1, characterized in that, The method further includes: After obtaining the initial input data through the task orchestration agent, the initial structure confidence index corresponding to the initial complex structure is calculated. Determine whether the initial structure confidence index is greater than a preset initial confidence threshold; If so, the initial complex structure is determined as the complex structure of the current round; The invocation structure preparation agent reads the complex structure of the current round and submits the complex structure of the current round to the binding site prediction agent.

3. The method according to claim 1, characterized in that, The step of invoking a residue-locking decision agent to determine whether the residues in the protein-protein conjugate to be optimized are locked includes: The residue-locking decision agent is invoked to read the binding site probability corresponding to each residue in multiple consecutive iterations. If the probability of the binding site of a residue is greater than or equal to a preset probability threshold in multiple consecutive iterations, then the residue is determined to be locked, and the residue is added to the locked residue set; or, Determine whether the residue is a pre-set permanently locked residue; if so, confirm that the residue is locked and add it to the set of locked residues; or... Determine whether the residue is outside the redesign scope; if so, determine that the residue is locked and add the residue to the locked residue set.

4. The method according to claim 3, characterized in that, The method further includes: Before adding the residue to the locked residue set, obtain the residue-level structural confidence of the residue; If the binding site probability of the residue is greater than or equal to the preset probability threshold, and the residue-level structural confidence level of the residue is greater than or equal to the preset structural confidence threshold, then the residue is identified as a candidate locked residue. The residue is added to the set of locked residues based on the candidate locked residues in multiple consecutive iterations.

5. The method according to claim 1, characterized in that, The steps of using a structural conditional sequence design model to design sequences of the residues in the set of freely designed sites to generate multiple candidate protein-conjugate sequences include: The complex structure corresponding to the current iteration, the protein conjugate sequence of the current iteration, the locked residue set, and the free design site set are input into the structure conditional sequence design model; By using the structural conditional sequence design model, while keeping the amino acids corresponding to the residues in the locked residue set unchanged, amino acid sampling or replacement is performed on the residues in the free design site set to generate multiple candidate protein-binding sequences.

6. The method according to claim 1, characterized in that, The step of invoking a structural evaluation agent to evaluate the candidate protein-binding sequence and the target protein sequence includes: The structure evaluation agent is invoked to input the candidate protein conjugate sequence and the target protein sequence into the evaluation model, and the corresponding candidate complex structure and the structure confidence index corresponding to the candidate complex structure are obtained through the evaluation model.

7. The method according to claim 1, characterized in that, The step of invoking the candidate selection agent to determine the seed for the next iteration from multiple candidate protein-protein conjugate sequences based on the structural confidence index includes: The candidate selection agent is invoked to rank multiple candidate protein-protein conjugate sequences according to the structural confidence index. The candidate protein conjugate sequences whose sorting results meet the preset selection criteria are determined as the seeds for the next round of iterations; The preset selection conditions include at least one of the following: the highest structural confidence index, the structural confidence index exceeding a preset candidate threshold, or the highest comprehensive score; the comprehensive score is obtained by weighting at least two of a plurality of indicators including the structural confidence index.

8. The method according to claim 1, characterized in that, The method further includes: The calling candidate selection agent feeds back the next round iteration seed to the task orchestration agent, so that the task orchestration agent schedules the functional agent to repeatedly execute the iterative optimization process based on the next round iteration seed, until the preset stopping condition is met and the optimization result is output. The preset stop condition includes at least one of the following: The confidence index of the optimal structure in the current round has reached the preset absolute threshold; The gain of the structural confidence index between two adjacent iterations is less than a preset gain threshold; The structural confidence index did not improve after a series of preset rounds; Reaching the maximum number of iterations; The number of free design sites in the set of free design sites is less than a preset number threshold; The number of locked residues in the locked residue set reaches a preset threshold. The coverage of the locked residues reaches a preset coverage threshold.

9. The method according to claim 1, characterized in that, The protein conjugate to be optimized is a short peptide conjugate; the preset redesign range includes all residues of the short peptide conjugate, and no permanently locked residues are set; or, The protein conjugate to be optimized is a nanobody; the preset redesign range is the complementarity-determining region (CDR) of the nanobody; the framework region residues of the nanobody are located outside the redesign range, and / or preset anchoring sites in the CDR are set as permanently locked residues; or... The protein conjugate to be optimized is a micro protein conjugate; the preset redesign range is the entire chain of the micro protein conjugate or a specified preset residue range; if the entire chain optimization mode is adopted, no permanently locked residues are set; if a specific backbone or functional site is retained, the corresponding residues are set as permanently locked residues.

10. An intelligent agent execution system, characterized in that, The intelligent agent execution system includes multiple functional intelligent agents to execute the method described in any one of claims 1 to 9.

Images