A multi-valent tumor mRNA vaccine design system and method for multi-dimensional combination optimization

Abstract

The application discloses a multi-valent tumor mRNA vaccine design system and method for multi-dimensional combination optimization, and the system comprises the following modules: an input module, which is used for receiving amino acid sequence information and design parameters of a plurality of target antigens; an MCTS sequence search module, which is used for generating candidate mRNA sequences in a multi-dimensional design space through an iteration mechanism with a Monte Carlo tree search algorithm as a core; a scoring function module, which is used for comprehensively evaluating the received candidate sequences in multiple dimensions by using an evaluation model, and feeding back the comprehensive score as a reward value to the MCTS sequence search module; and an output module, which is used for receiving and outputting the optimal mRNA complete sequence with the highest comprehensive score obtained by the MCTS sequence search module after the MCTS sequence search module reaches a preset iteration number. The technical scheme of the application realizes a global optimal balance of immunogenicity, mRNA stability and expression efficiency.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a multivalent tumor mRNA vaccine design system and method for multidimensional combinatorial optimization. Background Technology

[0002] To address the heterogeneity and immune evasion of tumor cells, multivalent mRNA vaccines can simultaneously encode multiple tumor-associated antigens or neoantigens, aiming to induce a broader and more durable immune response. Existing research indicates that linking multiple antigenic epitopes to the same mRNA molecule via linkers is a common strategy for achieving multivalent vaccines. For example, current technologies often employ universal linkers such as 2A peptides, furin restriction site linkers, and GS linkers to heuristically combine different antigen sequences for co-expression.

[0003] However, this traditional manual design method has the following technical problems and drawbacks: First, the design space is enormous, making global optimization difficult: taking 10 neoantigens as an example, there are 10! (approximately 3.6 million) possible permutations. If we also consider the type and number of linking peptides and the selection of synonymous codons at each site, the combination space becomes extremely large. Traditional methods rely on experience and trial and error, making it impossible to explore the optimal solution in such a multidimensional space.

[0004] Second, redundant linker peptides have negative effects: Traditional methods often require the insertion of multiple linker peptides to link multiple antigens. This not only significantly increases the length of the mRNA transcript, affecting the translation efficiency of the 3' end antigen, but may also lead to the formation of repetitive structures in the mRNA, increasing its instability, and even causing ribosome arrest, thereby reducing overall translation efficiency.

[0005] Third, balancing conflicting objectives is difficult: vaccine design needs to simultaneously meet multiple objectives, such as high immunogenicity, high mRNA stability, and high expression efficiency. However, different antigen sequences have different characteristics; highly immunogenic sequences may lead to complex mRNA structures, affecting their stability; while highly expressed sequences may have weaker immunogenicity. Current technologies lack effective means to coordinate the potential conflicts between these performance indicators.

[0006] Fourth, there is a lack of integrated and intelligent design tools: current multivalent vaccine construction methods mainly remain at the level of manual combination, failing to consider factors such as antigen sequencing, linker peptide optimization, and codon selection as a holistic system. There is an urgent need for a computational method that can integrate multiple evaluation models and perform intelligent search and optimization in the global design space to overcome existing design bottlenecks. Summary of the Invention

[0007] This invention aims to address the technical problems of existing multivalent tumor mRNA vaccine design, which relies on heuristic combinatorial design, fails to achieve global optimization in a multi-dimensional combinatorial space including antigen sequencing, linker peptide usage, and codon optimization, and struggles to balance multiple design objectives such as immunogenicity, mRNA stability, and expression efficiency. The purpose of this invention is to provide an integrated and intelligent multivalent mRNA vaccine design system and method to achieve efficient and stable rational design of multi-antigen mRNA sequences, thereby enhancing the vaccine's immune activation effect and anti-tumor capability.

[0008] To achieve the above objectives, the present invention provides the following solution: A multivalent tumor mRNA vaccine design system for multidimensional combinatorial optimization includes: The input module is used to receive amino acid sequence information and design parameters from multiple target antigens; The MCTS sequence search module is used to generate candidate mRNA sequences in a multidimensional design space based on the amino acid sequence information and design parameters, and with the Monte Carlo tree search algorithm as the core, through an iterative mechanism of "selection-expansion-simulation-backpropagation". The scoring function module is used to receive candidate mRNA sequences generated by the MCTS sequence search module, integrate an immunogenicity prediction model and an mRNA stability assessment model, perform a multi-dimensional comprehensive evaluation of the received candidate mRNA sequences, and feed back the comprehensive score as a reward value to the MCTS sequence search module. The output module is used to receive and output the optimal mRNA complete sequence with the highest comprehensive score obtained by the MCTS sequence search module after the MCTS sequence search module has reached a preset number of iterations.

[0009] Preferably, in the input module, the amino acid sequence information includes neoantigens derived from tumor sequencing data or known tumor-associated antigens; The design parameters include the linker peptide sequence, species, UTR sequence, and threshold for calling the tool.

[0010] Preferably, the MCTS sequence search module includes: The selection unit is used to select a partial sequence that meets the potential threshold from the child nodes of the current node, starting from the root node representing an empty sequence, based on the upper confidence interval algorithm, in order to balance the exploration of new paths and the utilization of known high-value paths. An extension unit is used to select an antigen from the remaining antigen pool at the selected node, connect it to either end of the current partial sequence, and generate a new partial sequence node. The simulation unit is used to start from a new partial sequence node, randomly link the remaining antigens one by one, and selectively insert predefined linker peptide sequences between adjacent antigens until a complete peptide sequence containing all input antigens is generated; wherein, the simulation unit further includes a codon optimization unit, used to perform synonymous codon optimization on the complete peptide sequence, thereby generating the corresponding candidate mRNA sequence.

[0011] The backpropagation unit is used to propagate the reward value fed back by the scoring function module back along the current search path, and update the access count and cumulative value statistics of all nodes on the path.

[0012] Preferably, the scoring function module includes: An immunogenicity scoring unit is used to obtain an immunogenicity score based on the immunogenicity prediction model. The mRNA stability scoring unit is used to obtain a stability score based on the mRNA stability assessment model and the optimized candidate mRNA sequence. The codon fitness index calculation unit is used to calculate the codon fitness index based on the optimized candidate mRNA sequence; The comprehensive scoring unit is used to perform a weighted comprehensive score on the immunogenicity score, stability score, and codon fitness index to obtain a comprehensive score as the reward value.

[0013] Preferably, in the scoring function module: The immunogenicity prediction model uses a machine learning model to quantitatively assess the potential of multi-antigen combinations to be cleaved and transported by the proteasome and to be bound by major histocompatibility complex molecules, in order to predict their ability to activate specific immune responses. The mRNA stability assessment model uses the DegScore assessment tool to predict degradation tendency based on mRNA sequence characteristics and quantify the stability of nucleic acid structure. Preferably, the codon optimization unit employs a two-stage optimization strategy: Phase 1: Based on the codon usage frequency table of the target species, multiple initial mRNA sequences are randomly sampled and generated, and the optimal initial sequence is selected through multi-index evaluation. The second stage involves using an mRNA degradation model to identify the key codon units with the highest susceptibility to hydrolysis in the initial sequence, enumerating all synonymous variants for each key codon unit and evaluating them using multiple metrics, iteratively selecting the optimal variant, and achieving progressive enhancement of the sequence.

[0014] This invention also provides a method for designing multivalent tumor mRNA vaccines with multidimensional combinatorial optimization, the system comprising: Obtain amino acid sequence information and design parameters for multiple target antigens; Based on the amino acid sequence information and design parameters, and using the Monte Carlo tree search algorithm as the core, candidate mRNA sequences are generated in a multidimensional design space through an iterative mechanism of "selection-expansion-simulation-backpropagation". Based on the candidate mRNA sequences, and integrating the immunogenicity prediction model and the mRNA stability assessment model, the received candidate mRNA sequences are comprehensively evaluated in multiple dimensions, and the comprehensive score is fed back to the Monte Carlo tree search algorithm as a reward value. After the Monte Carlo tree search algorithm reaches a preset number of iterations, it outputs the optimal complete mRNA sequence with the highest comprehensive score.

[0015] Preferred methods for generating candidate mRNA sequences in a multidimensional design space include: Starting from the root node representing an empty sequence, based on the upper confidence interval algorithm, a partial sequence that meets the potential threshold is selected from the child nodes of the current node for the next step of exploration, in order to balance the exploration of new paths and the utilization of known high-value paths; At the selected node, an antigen is chosen from the remaining antigen pool and connected to either end of the current partial sequence to generate a new partial sequence node; Starting from the new partial sequence node, the remaining antigens are randomly linked one by one, and predefined linker peptide sequences are selectively inserted between adjacent antigens until a complete peptide sequence containing all input antigens is generated. Synonymous codon optimization is then performed on the complete peptide sequence to generate the corresponding candidate mRNA sequence. The reward value is propagated backward along the current search path to update the visit count and cumulative value statistics for all nodes along the path.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention utilizes the MCTS algorithm and a joint scoring function to solve the technical problem that existing technologies cannot achieve multi-objective collaborative optimization in a multi-dimensional design space, achieving a globally optimal balance between immunogenicity, mRNA stability, and expression efficiency. By employing a multi-threaded parallelization strategy and a scalable algorithm architecture, it solves the technical problem of low computational efficiency or even inoperability in existing technologies when dealing with the design of long-sequence, high-cost vaccines, demonstrating excellent practicality and scalability. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a system operation flowchart of an embodiment of the present invention; Figure 2 This is a flowchart of a method according to an embodiment of the present invention; Figure 3 The diagram shows the effect analysis of the method proposed in this embodiment of the invention. Based on the 5000 candidate mRNA sequences obtained by the method, the correlation between multiple performance indicators was statistically analyzed. Among them, (a) is a schematic diagram of the performance of the proposed solution in the MHC class I neoantigen design task, (b) is a schematic diagram of the performance of the proposed solution in the MHC class II neoantigen design task, (c) is a distribution diagram of the relationship between the number of MHC class I neoantigen linking peptides and mRNA hydrolytic stability, (d) is a distribution diagram of the relationship between the number of MHC class II neoantigen linking peptides and mRNA hydrolytic stability, (e) is a distribution diagram of the relationship between the number of MHC class I neoantigen linking peptides and mRNA minimum free energy, (f) is a distribution diagram of the relationship between the number of MHC class II neoantigen linking peptides and mRNA minimum free energy, (g) is a distribution diagram of the relationship between ... Figure 4 The figures show the performance verification results of the multi-threaded parallelization strategy and scalable algorithm architecture of the present invention. (a) and (b) compare the sequence performance metrics of parallel search (labeled p) and non-parallel search under the same task, with the numbers in parentheses representing the total number of sequences searched under the corresponding conditions. (c) shows the running time comparison of the two sets of tasks. (d) is a schematic diagram of the immune score when 8, 16, and 24 neoantigens are input in the MHC class II neoantigen design task, respectively. (e) is a schematic diagram of the number of linker peptides when 8, 16, and 24 neoantigens are input in the MHC class II neoantigen design task, respectively. (f) is a schematic diagram of the degradation score when 8, 16, and 24 neoantigens are input in the MHC class II neoantigen design task, respectively. (g) is a schematic diagram of the CAI when 8, 16, and 24 neoantigens are input in the MHC class II neoantigen design task, respectively. (h) is a schematic diagram of the degradation score when 8, 16, and 24 neoantigens are input in the MHC class II neoantigen design task, respectively. In the MHC Class II neoantigen design task, the GC content diagrams for inputting 8, 16, and 24 neoantigens respectively are shown; (i) is the computation time corresponding to two example configurations (NeoWeaver-B and NeoWeaver-V) under the MHC Class II neoantigen design task; Figure 5 This is a flowchart of the codon optimization algorithm according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments 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, and 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.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Example 1: like Figure 1 As shown, a multivalent tumor mRNA vaccine design system for multidimensional combinatorial optimization includes: an input module, an MCTS sequence search module, a scoring function module, and an output module.

[0022] An input module is used to receive amino acid sequence information and design parameters for multiple target antigens; in a further embodiment, the amino acid sequence information in the input module includes novel antigens derived from tumor sequencing data or known tumor-associated antigens; the design parameters include linker peptide sequences, species, UTR sequences, and thresholds for calling tools.

[0023] The MCTS sequence search module is used to generate candidate mRNA sequences in a multidimensional design space based on amino acid sequence information and design parameters, using the Monte Carlo tree search algorithm as its core, and through an iterative mechanism of "selection-expansion-simulation-backpropagation". A further implementation of the MCTS sequence search module includes: The selection unit, starting from the root node representing an empty sequence, selects a partial sequence from the child nodes of the current node that meets a potential threshold for further exploration, based on the upper confidence interval algorithm, to balance exploring new paths and utilizing known high-value paths. Specifically, starting from the root node (empty sequence), for nodes that are fully expanded and have visit records, the optimal child node is selected according to the UCT formula: in, Represents a node The average return (i.e., total reward divided by the number of visits). For nodes Number of visits, The total number of simulations for the parent node. To explore the constant. The first criterion favors nodes with higher empirical rewards, reflecting exploitation; the second favors nodes with fewer visits, reflecting exploration. This is achieved by adjusting the constant. A balance can be struck between exploration and utilization.

[0024] The expansion unit is used to select an antigen from the remaining antigen pool at the selected node, connect it to either end of the current partial sequence, and generate a new partial sequence node. The simulation unit, starting from a new partial sequence node, randomly links the remaining antigens one by one, selectively inserting predefined linker peptide sequences between adjacent antigens until a complete peptide sequence containing all input antigens is generated. The simulation unit also includes a codon optimization unit for performing synonymous codon optimization on the complete peptide sequence to generate corresponding candidate mRNA sequences. Figure 5 As shown. A further implementation method involves NeoWeaver-B using its self-developed SynCoop algorithm for codon optimization, employing a two-stage optimization strategy: Phase 1: Based on the codon usage frequency table of the target species, multiple initial mRNA sequences are randomly sampled and generated, and the optimal initial sequence is selected through evaluation using multiple indicators (including CAI, DegScore, GC content, etc.). The second stage involves using an mRNA degradation model to identify the key codon units with the highest susceptibility to hydrolysis in the initial sequence, enumerating all synonymous variants for each key codon unit and evaluating them using multiple metrics, iteratively selecting the optimal variant, and achieving progressive enhancement of the sequence.

[0025] The Backpropagation unit is used to propagate the reward value fed back by the scoring function module back along the current search path, updating the visit count and cumulative value statistics of all nodes on the path.

[0026] The scoring function module receives candidate mRNA sequences generated by the MCTS sequence search module, integrates an immunogenicity prediction model and an mRNA stability assessment model, performs a multi-dimensional comprehensive evaluation of the received candidate mRNA sequences, and feeds back the comprehensive score as a reward value to the MCTS sequence search module. A further implementation method is that the scoring function module includes: The Immunogenicity Score unit is used to obtain an immunogenicity score based on the aforementioned immunogenicity prediction model. Specifically, for MHC class I antigens, NetChop and Pepsickle are first used to predict proteasome cleavage sites, and then MHCflurry and NetMHCpan are used to predict the binding affinity of the resulting peptides to MHC class I alleles, and the immunogenicity score is calculated by combining these methods. For MHC class II antigens, ProsperousPlus is first used to predict enzyme cleavage, and then NetMHCIIpan is used to predict the binding affinity of the resulting peptides to MHC class II alleles, and the immunogenicity score is calculated by combining these methods.

[0027] The mRNA stability score (DegScore) unit is used to obtain a stability score based on the mRNA stability assessment model and the optimized candidate mRNA sequence. The Codon Fitness Index (CAI) calculation unit is used to calculate the codon fitness index based on the optimized candidate mRNA sequence. A further implementation method is found in the scoring function module: An immunogenicity prediction model uses machine learning to quantitatively assess the potential of multi-antigen combinations to be cleaved and transported by the proteasome and to be bound by major histocompatibility complex molecules, in order to predict their ability to activate specific immune responses. The mRNA stability assessment model uses the DegScore assessment tool to predict degradation tendency based on mRNA sequence characteristics and quantify the stability of nucleic acid structures. The comprehensive scoring unit is used to perform a weighted comprehensive score on the immunogenicity score, stability score, and codon fitness index to obtain a comprehensive score as the reward value.

[0028] Linker peptide penalty term: A penalty term is introduced when the number of linker peptides inserted into the sequence exceeds a preset threshold. , In this embodiment, the weights are set differently based on the number of antigen peptides to be linked and the MHC type. Represents immunogenicity weight. Represents the codon fitness weight. This represents the stability weight.

[0029] The output module receives and outputs the optimal complete mRNA sequence with the highest overall score obtained by the MCTS sequence search module after the MCTS sequence search module has reached a preset number of iterations. The optimal sequence includes optimized antigen arrangement, linker peptide usage strategy, and synonymous codon combination.

[0030] Example 2: Construction and Operation of the NeoWeaver Multivalent Tumor mRNA Vaccine Design System 1. Experimental Materials Data source: Amino acid sequences of multiple candidate neoantigens were identified and screened from publicly available databases or user-provided tumor sequencing data. In this embodiment, 8 experimentally validated MHC class I neoantigens and 24 MHC class II neoantigens derived from the CT26 cell line were used as the test dataset.

[0031] Computing platform: Linux server equipped with multi-core CPU and high memory (such as 256GB RAM).

[0032] 2. Experimental Methods 2.1 NeoWeaver System Framework Setup According to the technical solution of this invention, a NeoWeaver multivalent mRNA vaccine design system is constructed. This system mainly includes an input module, an MCTS sequence search module, a scoring function module, and an output module. All modules are implemented using the Python 3.8 programming language.

[0033] 2.2 Preparation of Test Data and Parameters 2.2.1 Neoantigen peptide dataset Neoantigen peptides of mouse MHC class I and II molecules, validated for immunogenicity using the ELISpot assay, were selected. All neoantigens were derived from the CT26 tumor model of the BALB / c mouse strain, a widely used inbred strain in cancer immunology research. Peptide sequences were collected from two published studies. Eight MHC class I and 24 MHC class II restricted neoantigen peptides were used for computational evaluation.

[0034] 2.2.2 Non-translated region sequences Both the 5' untranslated region and the 3' untranslated region use the UTR sequence of the human β-globin gene, a combination of sequences widely used in mRNA vaccine design.

[0035] 2.2.3 Linking peptide sequence All linker peptides were constructed using the standard flexible linker peptide motif GGGGSGGGGS, a glycine-serine enriched sequence that has been widely used as a linker in epitope fusion constructs.

[0036] 2.2.4 Species Codon Preference BALB / c mice were used as the target species, and the codon usage frequency table was obtained from the KAZUSA database for subsequent synonymous codon optimization.

[0037] 2.3 MCTS Sequence Search Process We will use a model performance test experiment as an example to introduce the search process of MCTS. Eight new MHC class I and MHC class II antigens were selected respectively. The exploration constant C of the MCTS algorithm was set to 5, and the preset number of iterations was 5000 rounds. The following steps were performed in each iteration: Selection: Starting from the root node (empty sequence), for nodes that are fully expanded and have visited records, select the optimal child node according to the UCT formula: in, Represents a node The average return (i.e., total reward divided by the number of visits). For nodes Number of visits, The total number of simulations for the parent node. To explore the constant. The first criterion favors nodes with higher empirical rewards, reflecting exploitation; the second favors nodes with fewer visits, reflecting exploration. This is achieved by adjusting the constant. A balance can be struck between exploration and utilization.

[0038] The selection step is performed only if the current node has child nodes and all child nodes have been visited at least once. In this problem, the root node initially corresponds to an empty solution and has no visited child nodes. Therefore, the algorithm skips the selection step in the initial stage and proceeds directly to the expansion step.

[0039] Extension: When encountering a node that is not fully expanded, a new antigen is randomly selected from the remaining antigen pool, and a linking strategy (linking peptide or none) is randomly selected to add it to the end of the current partial sequence to form a new child node.

[0040] Simulation: Starting from the new node, the remaining antigens are randomly sorted and connected (randomly selecting the connection strategy) to generate a complete peptide chain sequence containing all 8 antigens.

[0041] Scoring: The scoring function module is called to comprehensively evaluate the complete sequence generated by the simulation.

[0042] The immunogenicity score is calculated as follows: For MHC class I antigens, NetChop and Pepsickle are first used to predict proteasome cleavage sites, and then MHCflurry and NetMHCpan are used to predict the binding affinity of the resulting peptides to MHC class I alleles. For MHC class II antigens, ProsperousPlus is first used to predict enzyme cleavage, and then NetMHCIIpan is used to predict the binding affinity of the resulting peptides to MHC class II alleles. The immunogenicity score is then calculated as follows:

[0043] mRNA stability score (DegScore): After codon optimization of the sequence, the DegScore model is used to predict the degradation tendency of the optimized mRNA sequence.

[0044] Codon fitness index (CAI): Calculates the codon fitness index of the optimized sequence to measure its compatibility with the host translation mechanism.

[0045] Linker peptide penalty term: A penalty term is introduced when the number of linker peptides inserted into the sequence exceeds a preset threshold.

[0046] In this embodiment, the weights are set differently based on the number of antigen peptides to be linked and the MHC type. For the MHC type, the immunogenicity weight is set to... = 10 / 8 = 1.25; For MHC type II, set to = 15 / 8 ≈ 1.85; the codon fitness weight is uniformly set to = 100, stability weight set to = -0.2, the above weighting is set to balance the relative contributions of immunogenicity score, stability score and codon fitness score in the overall evaluation.

[0047] Backpropagation: The total score obtained in this simulation is backpropagated along all the nodes visited in this iteration, updating the total reward value and number of visits for each node.

[0048] 2.4 Codon optimization algorithm (SynCoop), such as Figure 5 As shown.

[0049] In this embodiment, NeoWeaver-B uses its self-developed SynCoop algorithm for codon optimization: Phase 1 (Initial Sequence Generation): Based on the codon usage frequency table of the target species, n initial mRNA sequences (n=80) are randomly sampled and generated. All candidate sequences are scored using a multi-index evaluation framework (including CAI, DegScore, GC content, etc.), and the 8 sequences with the highest scores are selected as the starting point for initial optimization.

[0050] The second stage (targeted optimization) uses an mRNA degradation model to identify the nucleotide positions in the initial sequence with the highest susceptibility to hydrolysis, locating the corresponding m=30 key codon units. For these 30 target codons, all synonymous variants are enumerated one by one and scored using the same multi-index evaluation system. Optimization is performed iteratively, with the best-performing variant selected as the input for the next round at each step, achieving progressive sequence enhancement.

[0051] 2.5 Control Experiment Setup To verify the effectiveness of this invention, the following control group was set up: NeoWeaver-B group (in this invention embodiment): The MCTS algorithm is combined with the self-developed SynCoop codon optimization algorithm. The process is run for 5000 rounds according to the above procedure to select the 5 non-redundant sequences with the highest total score.

[0052] The NeoWeaver-V group (a variant of this invention): A variant of NeoWeaver-B that combines MCTS search with the mainstream structure-driven algorithm LinearDesign. After generating the complete peptide chain in the simulation step, LinearDesign is called for codon optimization and stability design. It also runs for 5000 rounds, selecting the 5 non-redundant sequences with the highest total scores.

[0053] Baseline group (traditional method comparison): Simulates the traditional modular vaccine design process. Random antigen sequencing is used, and a universal linker peptide is inserted between all antigens, followed by optimization using LinearDesign. Five sequences are randomly generated for comparison.

[0054] In summary, this invention models multivalent mRNA vaccine design as a combinatorial optimization problem and introduces the Monte Carlo Tree Search (MCTS) algorithm for global solution, establishing the overall inventive concept of "multi-dimensional evaluation driving intelligent search, and joint optimization achieving performance balance." To systematically verify the technical effects of this invention, we used experimentally validated MHCI and MHC II neoantigens derived from CT26 as test objects, and conducted a comprehensive comparative analysis of the proposed NeoWeaver (containing two optimization strategies, NeoWeaver-B and NeoWeaver-V), with traditional modular design strategies. NeoWeaver-B consists of MCTS + the autonomous codon optimization algorithm SynCoop, while NeoWeaver-V consists of MCTS + the mainstream structure-driven algorithm LinearDesign. Traditional strategies involve random antigen sorting + linker peptide insertion + LinearDesign optimization. Experimental results fully demonstrate the significant progress of this invention in solving existing technical problems. Specific technical effects are as follows: 1. This invention utilizes the MCTS algorithm and a joint scoring function to solve the technical problem that existing technologies cannot achieve multi-objective collaborative optimization in a multi-dimensional design space, thus achieving a globally optimal balance between immunogenicity, mRNA stability, and expression efficiency. For example... Figure 3 As shown.

[0055] Leading Immunogenicity: In the MHC class I neoantigen test, the immunogenicity scores of the sequences generated by NeoWeaver-B and NeoWeaver-V were significantly higher than those of the baseline control group. This indicates that the optimization process of this invention does not sacrifice core immune activation capabilities, but rather synergistically optimizes other performance indicators while ensuring high immunogenicity.

[0056] Significantly improved expression efficiency: NeoWeaver-B performed best in the codon fit index (CAI), a key indicator of translation efficiency, with a CAI value significantly higher than NeoWeaver-V and Baseline. This directly demonstrates that the codon optimization algorithm (SynCoop) independently developed in this invention can more effectively adapt to the host translation mechanism, laying a solid foundation for the efficient expression and presentation of subsequent antigens.

[0057] Rational regulation of mRNA structural stability: Thermodynamic stability (MFE): Data shows that as the number of linkers increases, the minimum free energy (MFE) of mRNA decreases, indicating enhanced thermodynamic stability. The sequence generated by NeoWeaver-V has the lowest MFE, which aligns with LinearDesign's MFE-driven optimization logic.

[0058] Hydrolytic stability (DegScore): However, a decrease in MFE is not always beneficial. Analysis shows that excessively increasing the number of linkers can actually lead to a decrease in hydrolytic stability (DegScore), and the GC content may deviate from the ideal range (this problem exists in the Baseline group). NeoWeaver-V achieved the lowest degradation risk (DegScore), while NeoWeaver-B controlled the GC content at the ideal level of around 0.5, avoiding synthesis difficulties and decreased translation efficiency caused by excessively high or low GC content, each with its own advantages.

[0059] As previously mentioned, while increasing the number of linkers can reduce MFE (enhanced thermodynamic stability), it also impairs hydrolytic stability (DegScore). This invention uses MCTS search to automatically find the optimal balance point in this trade-off. Analysis shows that NeoWeaver-B tends to generate sequences with no or very few linkers, while NeoWeaver-V uses a wider distribution of linkers (1 to 7), a strategy adopted to pursue extremely low MFE. Notably, even in NeoWeaver-V, its high-scoring sequences avoid full linker splicing, indicating that the MCTS framework of this invention can effectively suppress the excessive pursuit of a single indicator (such as MFE), ensuring that the final output sequence also possesses biological functionality.

[0060] This invention, through a joint scoring function, can precisely control performance indicators across different dimensions. NeoWeaver-B achieves an optimal balance in immunogenicity, CAI, and linker control, resulting in shorter and more efficient vaccine sequences; NeoWeaver-V, on the other hand, demonstrates superior RNA stability. Both significantly outperform the baseline group, which performed poorly across all core indicators, highlighting the necessity and advancement of this invention's "guided search" and "multi-objective joint optimization" strategies.

[0061] 2. This invention utilizes a multi-threaded parallelization strategy and a scalable algorithm architecture to solve the technical problem of low computational efficiency or even inoperability in existing technologies when dealing with the design of long sequences and high-cost vaccines, demonstrating excellent practicality and scalability. For example... Figure 4 As shown.

[0062] To address the design requirements for multiple antigens (such as 16, 24, or more) that may arise in clinical applications, this invention introduces a leaf node parallel strategy and a virtual loss mechanism into the MCTS algorithm, and implements differentiated parallel designs for different types of computational tasks, achieving significant results: Search efficiency has been greatly improved: In the testing of eight new MHC class I antigens, the use of eight parallel threads improved the overall search speed by about three times, greatly shortening the vaccine design cycle and ensuring the timeliness of personalized treatment.

[0063] Excellent ability to process long sequences: When the number of antigen inputs increased from 8 to 16, both NeoWeaver-B and NeoWeaver-V operated stably. NeoWeaver-B maintained its lead in immunogenicity and CAI, and demonstrated strategic flexibility (shifting from a preference for no linker to a moderate introduction of linkers in exchange for better overall performance).

[0064] When the number of inputs increased to 24, NeoWeaver-V and Baseline, which rely on LinearDesign, failed to run due to their near-quadratic algorithm complexity, leading to exhaustion of computational resources. In contrast, NeoWeaver-B still demonstrated excellent stability and scalability: its immunogenicity score increased approximately linearly with the number of inputs, the average linker usage was kept below 5, and both CAI and GC content remained within a reasonable range.

[0065] The computation time increased approximately linearly: from 8 to 16 antigens, NeoWeaver-V's computation time surged from 3.4 hours to 33.1 hours, while NeoWeaver-B's time only increased from 4.4 hours to 9.5 hours, showing a healthy, approximately linear increase. This fully demonstrates the architectural advantages of the core algorithm of this invention, which can efficiently utilize hardware resources to process longer sequences, paving the way for designing high-value vaccines covering more targets.

[0066] In summary, this invention, through the deep coupling of the MCTS intelligent search framework and a multi-dimensional joint scoring function, successfully solves the three core challenges of existing technologies: "inability to achieve global optimization," "redundant linkers impairing function," and "low efficiency in long sequence design." This invention not only designs mRNA sequences with stronger immunogenicity, more stable structure, and higher expression efficiency, but also demonstrates unparalleled efficiency and scalability in handling complex and costly vaccine design tasks. It provides a powerful computational engine for the clinical translation of personalized tumor mRNA vaccines, possessing significant inventiveness and progressiveness.

[0067] Example 3: like Figure 2 As shown, the present invention also provides a method for designing multivalent tumor mRNA vaccines for multidimensional combinatorial optimization, using the system of Example 1, comprising: Obtain amino acid sequence information and design parameters for multiple target antigens; Based on amino acid sequence information and design parameters, and with the Monte Carlo tree search algorithm as the core, candidate mRNA sequences are generated in a multidimensional design space through an iterative mechanism of "selection-expansion-simulation-backpropagation". Based on candidate mRNA sequences, and integrating immunogenicity prediction models and mRNA stability assessment models, the received candidate mRNA sequences are comprehensively evaluated from multiple dimensions, and the comprehensive score is fed back as a reward value to the Monte Carlo tree search algorithm. After the Monte Carlo tree search algorithm reaches the preset number of iterations, it outputs the optimal complete mRNA sequence with the highest comprehensive score.

[0068] A further implementation method for generating candidate mRNA sequences in a multidimensional design space includes: Starting from the root node representing an empty sequence, based on the upper confidence interval algorithm, a partial sequence that meets the potential threshold is selected from the child nodes of the current node for the next step of exploration, in order to balance the exploration of new paths and the utilization of known high-value paths; At the selected node, an antigen is chosen from the remaining antigen pool and connected to either end of the current partial sequence to generate a new partial sequence node; Starting from the new partial sequence node, the remaining antigens are randomly linked one by one, and predefined linker peptide sequences are selectively inserted between adjacent antigens until a complete peptide sequence containing all input antigens is generated. Synonymous codon optimization is then performed on the complete peptide sequence to generate the corresponding candidate mRNA sequence.

[0069] The reward value is propagated backward along the current search path, updating the visit count and cumulative value statistics for all nodes along the path.

[0070] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A multivalent tumor mRNA vaccine design system for multidimensional combinatorial optimization, characterized in that, include: The input module is used to receive amino acid sequence information and design parameters from multiple target antigens; The MCTS sequence search module is used to generate candidate mRNA sequences in a multidimensional design space based on the amino acid sequence information and design parameters, and with the Monte Carlo tree search algorithm as the core, through an iterative mechanism of "selection-expansion-simulation-backpropagation". The scoring function module is used to receive candidate mRNA sequences generated by the MCTS sequence search module, integrate an immunogenicity prediction model and an mRNA stability assessment model, perform a multi-dimensional comprehensive evaluation of the received candidate mRNA sequences, and feed back the comprehensive score as a reward value to the MCTS sequence search module. The output module is used to receive and output the optimal mRNA complete sequence with the highest comprehensive score obtained by the MCTS sequence search module after the MCTS sequence search module has reached a preset number of iterations.

2. The system according to claim 1, characterized in that, In the input module, the amino acid sequence information includes neoantigens derived from tumor sequencing data or known tumor-associated antigens; The design parameters include the linker peptide sequence, species, UTR sequence, and threshold for calling the tool.

3. The system according to claim 1, characterized in that, The MCTS sequence search module includes: The selection unit is used to select a partial sequence that meets the potential threshold from the child nodes of the current node, starting from the root node representing an empty sequence, based on the upper confidence interval algorithm, in order to balance the exploration of new paths and the utilization of known high-value paths. An extension unit is used to select an antigen from the remaining antigen pool at the selected node, connect it to either end of the current partial sequence, and generate a new partial sequence node. The simulation unit is used to start from a new partial sequence node, randomly link the remaining antigens one by one, and selectively insert predefined linker peptide sequences between adjacent antigens until a complete peptide sequence containing all input antigens is generated; wherein, the simulation unit further includes a codon optimization unit, used to perform synonymous codon optimization on the complete peptide sequence, thereby generating the corresponding candidate mRNA sequence; The backpropagation unit is used to propagate the reward value fed back by the scoring function module back along the current search path, and update the access count and cumulative value statistics of all nodes on the path.

4. The system according to claim 3, characterized in that, The scoring function module includes: An immunogenicity scoring unit is used to obtain an immunogenicity score based on the immunogenicity prediction model. The mRNA stability scoring unit is used to obtain a stability score based on the mRNA stability assessment model and the optimized candidate mRNA sequence. The codon fitness index calculation unit is used to calculate the codon fitness index based on the optimized candidate mRNA sequence; The comprehensive scoring unit is used to perform a weighted comprehensive score on the immunogenicity score, stability score, and codon fitness index to obtain a comprehensive score as the reward value.

5. The system according to claim 1, characterized in that, In the scoring function module: The immunogenicity prediction model uses a machine learning model to quantitatively assess the potential of multi-antigen combinations to be cleaved and transported by the proteasome and to be bound by major histocompatibility complex molecules, in order to predict their ability to activate specific immune responses. The mRNA stability assessment model uses the DegScore assessment tool to predict degradation tendency based on mRNA sequence characteristics and quantify the stability of nucleic acid structures.

6. The system according to claim 3, characterized in that, The codon optimization unit employs a two-stage optimization strategy: Phase 1: Based on the codon usage frequency table of the target species, multiple initial mRNA sequences are randomly sampled and generated, and the optimal initial sequence is selected through multi-index evaluation. The second stage involves using an mRNA degradation model to identify the key codon units with the highest susceptibility to hydrolysis in the initial sequence, enumerating all synonymous variants for each key codon unit and evaluating them using multiple metrics, iteratively selecting the optimal variant, and achieving progressive enhancement of the sequence.

7. A method for designing multivalent tumor mRNA vaccines for multidimensional combinatorial optimization, using the system described in any one of claims 1-6, characterized in that, include: Obtain amino acid sequence information and design parameters for multiple target antigens; Based on the amino acid sequence information and design parameters, and using the Monte Carlo tree search algorithm as the core, candidate mRNA sequences are generated in a multidimensional design space through an iterative mechanism of "selection-expansion-simulation-backpropagation". Based on the candidate mRNA sequences, and integrating the immunogenicity prediction model and the mRNA stability assessment model, the received candidate mRNA sequences are comprehensively evaluated in multiple dimensions, and the comprehensive score is fed back to the Monte Carlo tree search algorithm as a reward value. After the Monte Carlo tree search algorithm reaches a preset number of iterations, it outputs the optimal complete mRNA sequence with the highest comprehensive score.

8. The method according to claim 7, characterized in that, Methods for generating candidate mRNA sequences in a multidimensional design space include: Starting from the root node representing an empty sequence, based on the upper confidence interval algorithm, a partial sequence that meets the potential threshold is selected from the child nodes of the current node for the next step of exploration, in order to balance the exploration of new paths and the utilization of known high-value paths; At the selected node, an antigen is chosen from the remaining antigen pool and connected to either end of the current partial sequence to generate a new partial sequence node; Starting from the new partial sequence node, the remaining antigens are randomly linked one by one, and predefined linker peptide sequences are selectively inserted between adjacent antigens until a complete peptide sequence containing all input antigens is generated. Synonymous codon optimization is then performed on the complete peptide sequence to generate the corresponding candidate mRNA sequence. The reward value is propagated backward along the current search path to update the visit count and cumulative value statistics for all nodes along the path.

Images