Codon sequence design method and apparatus based on large multi-modal model

Abstract

The application provides a protein coding codon sequence design method and device based on a large multimodal model, which comprehensively utilizes information such as amino acid sequences, spatial folding, mRNA quality and the like of a protein, and realizes high-quality, controllable and evaluative codon sequence design. The application breaks through the dependence of a traditional codon optimization strategy on a reference gene template, and converts a codon optimization problem into a codon design task based on protein information. By comprehensively considering information such as spatial folding, context of an amino acid sequence, mRNA quality and the like of a protein, the model can generate a codon sequence that is more in line with codon usage bias of a specific species, reasonable in structure and efficient in expression, thereby providing a new idea and technical support for safe, stable, efficient and controllable protein expression.

Description

Technical Field

[0001] This invention relates to the field of protein-coding codon design technology, and in particular, to a method and apparatus for codon sequence design based on a large-scale multimodal model. Background Technology

[0002] Proteins play crucial roles as participants and executors in life activities, widely involved in various life processes such as cell structure building, metabolism, signal transduction, and immune responses. Achieving controllable protein production has broad value in multiple application areas, including recombinant protein drug development, vaccine design, gene therapy, and synthetic biology research, and is of great significance for the development of cutting-edge biotechnology and the promotion of related industries.

[0003] The expression and production of proteins in cells is essentially the process by which ribosomes synthesize peptide chains based on genetic information carried by messenger ribonucleic acid (mRNA). In biology, a codon is a set of three consecutive bases in mRNA or the corresponding deoxyribonucleic acid (DNA) segment that determines one amino acid. In the translation process from gene to mRNA and then to protein, each codon encodes a specific amino acid. There are four types of bases in nature: adenine (A), cytosine (C), guanine (G), and thymine (T). These three bases can combine to form 64 different codons, corresponding to only 20 common amino acids, plus 21 signals for terminating peptide chain production. Therefore, there is redundancy in the coding relationships between codons and amino acids. Except for methionine and tryptophan, the other 18 basic amino acids are encoded by 2 to 6 codons. Different codons encoding the same amino acid are synonymous codons.

[0004] While the mapping from codons to amino acids is deterministic, the mapping from amino acids to codons is indeterminate for the 18 basic amino acids that have synonymous codons. To date, humans have not yet discovered the patterns by which organisms determine how to invoke synonymous codons during evolution.

[0005] Some known regular factors include: 1) Different species have different preferences for codon selection, and the corresponding codon optimization method is to increase the overall codon preference level of mRNA, that is, the codon optimization strategy of maximizing CAI (Codon Adaptation Index); 2) The GC content in mRNA is also different in different species. Too high a GC content will lead to the formation of complex secondary structures in mRNA, while too low a GC content will slow down peptide chain synthesis; 3) High-speed expression codons and low-speed expression codons in mRNA should appear in combination to increase the likelihood of obtaining correctly folded proteins.

[0006] However, some of the aforementioned principles are too general and lack operability, while others perform poorly in practice and cannot solve the problem of efficient protein expression. In fact, current traditional codon optimization algorithms have significant limitations: First, these methods primarily rely on the frequency of codon usage in the host cell, expecting high-frequency codons to lead to rapid peptide synthesis, while ignoring the context of the gene sequence; second, existing research has shown that current mainstream codon optimization strategies may interfere with protein folding conformation and stability, thereby altering their biological functions and even introducing safety risks to the production of recombinant protein drugs.

[0007] In recent years, with the development of artificial intelligence technology, attempts have emerged to introduce deep learning methods into the field of codon optimization. Specifically, this involves constructing neural network models to learn the mapping relationship between gene sequences and expression levels, thereby achieving intelligent optimization of codon sequences. Representative works include models using recurrent neural networks (RNNs), convolutional neural networks (CNNs), and Transformer architectures. However, these methods still use linear sequences as input and fail to fully incorporate the spatial structure and semantic context information of proteins, thus failing to address the disconnect between peptide expression and spatial folding. Furthermore, due to the limited availability and diverse sources of specific experimental data on protein expression levels, coupled with significant differences in methodologies, truly effective deep learning models cannot be trained. Summary of the Invention

[0008] The purpose of this invention is to provide a method and apparatus for designing codon sequences based on a large-scale multimodal model, so as to solve the limitations of traditional codon optimization algorithms.

[0009] To achieve the above objectives, the present invention provides a codon sequence design apparatus based on a large-scale multimodal model, comprising the following four parts: (A) Multimodal information encoding module: including protein three-dimensional structure acquisition layer and amino acid sequence acquisition layer. The former is directly input into the protein three-dimensional structure encoding layer, or the amino acid sequence is input into the protein three-dimensional structure prediction layer when the experimental structure cannot be obtained, and the latter is directly input into the amino acid sequence encoding layer. (B) Large-scale multimodal model-driven generation module: including a codon sequence generator and a temperature sampling layer. The protein three-dimensional structure coding layer, amino acid sequence coding layer and sampling temperature are input into the codon sequence generator, which generates multiple candidate sequences through the temperature sampling layer. (C) Sequence scoring and screening module: The multiple candidate sequences generated by the generation module are scored using multidimensional weighted indicators, including various mRNA quality assessment indicators. High-scoring codon sequences are selected based on the scoring results.

[0010] The present invention also provides a method for designing codon sequences based on a large-scale multimodal model using the aforementioned device, the specific steps of which are as follows: S1. Determine whether the protein's three-dimensional structure can be obtained from a public database based on the amino acid sequence: if the acquisition fails, obtain the protein's three-dimensional structure code through protein three-dimensional structure prediction; if the acquisition is successful, obtain the protein's three-dimensional structure code directly from the database. S2. Input the protein's three-dimensional structure encoding, amino acid sequence encoding, and sampling temperature into the codon sequence generator to generate codon sequences through a large-scale multimodal model; S3. Codon sequences are designed by using sequence scoring and screening based on multiple mRNA quality assessment indicators to obtain high-scoring codon sequences.

[0011] As one of the preferred technical solutions, in step S2, the codon sequence is generated through a label loss function and model training.

[0012] As a further preferred technical solution, the label loss function is the cross-entropy (CE) loss commonly used in language models. And the sum of any possible additional loss functions (such as the smoothing loss function).

[0013] As one of the further preferred technical solutions, let the batch size be B, the total number of password subcategories be C, and the sample length be L; Represented as: in, It is the one-hot encoding of the target tag. This represents the model output.

[0014] The overall loss function is the sum of all losses: As one of the further preferred technical solutions, the dataset is randomly divided into training set, validation set and test set during model training.

[0015] As one of the preferred technical solutions, in step S3, the sequence scoring index is: sorting and screening of multidimensional weighted indexes.

[0016] As one of the preferred technical solutions, in step S3, the sequence screening method includes, but is not limited to: CAI, GC content control, etc.

[0017] The present invention has the following beneficial effects: This invention provides a method and apparatus for codon sequence design based on a large-scale multimodal model. It comprehensively utilizes information such as protein amino acid sequence, spatial folding, and mRNA quality to achieve high-quality, controllable, and evaluable codon sequence design. This invention overcomes the dependence of traditional codon optimization strategies on reference gene templates, transforming the codon optimization problem into a codon design task based on protein information. By comprehensively considering information such as protein spatial folding structure, amino acid sequence context, and mRNA quality, the model can generate codon sequences that are more in line with the codon usage preferences of specific species, have reasonable structures, and are highly efficient in expression. This provides new ideas and technical support for safe, stable, efficient, and controllable protein expression.

[0018] Compared with previous deep learning-based codon optimization models, this invention has the following three major advantages.

[0019] First, this invention allows for the design of codon sequences directly from amino acid sequences, eliminating the need for template codon sequences and enabling precise end-to-end design and flexible adaptation, thus freeing the user from dependence on starting reference template codon sequences.

[0020] Secondly, by comprehensively considering information such as the amino acid sequence context of proteins and protein spatial folding, this invention enables models to better perceive the relationship between codon usage and various influencing factors, thereby improving the success rate of generating codon sequences.

[0021] Finally, the model of this invention combines a temperature sampling mechanism and has a unique confidence screening module, which ensures the diversity and high quality of the generated sequences.

[0022] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0023] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic diagram of a deep codon sequence generation method based on a large-scale multimodal model.

[0024] Figure 2 Schematic diagram of multimodal information encoding module.

[0025] Figure 3 A schematic diagram of a codon sequence generation module based on a large-scale multimodal model.

[0026] Figure 4This study compared the expression of human dihydrolactoxin using codon sequences designed in this method with other codon sequences using SAS-PAGE. The results are as follows: 1. Original sequence - before induction; 2. Original sequence - after induction; 3. Original sequence - supernatant; 4. Original sequence - precipitate; 5. Original sequence - nickel affinity column flow-through; 6. Original sequence - nickel affinity column washing; 7. Original sequence - nickel affinity column elution; 8. Optimized sequence using traditional algorithm - before induction; 9. Optimized sequence using traditional algorithm - after induction; 10. Optimized sequence using traditional algorithm - supernatant; 11. Optimized sequence using traditional algorithm - precipitate; 12. Optimized sequence using traditional algorithm - nickel affinity column flow-through; 13. Optimized sequence using traditional algorithm - nickel affinity column washing; 14. Optimized sequence using traditional algorithm - nickel affinity column elution; 15. Sequence 1 designed in this method - before induction; 16. 17. Supernatant of Sequence 1 of this method; 8. Precipitate of Sequence 1 of this method; 19. Flow-through of nickel affinity column of Sequence 1 of this method; 20. Washing of impurities in nickel affinity column of Sequence 1 of this method; 21. Elution of nickel affinity column of Sequence 1 of this method; 22. Before induction of Sequence 2 of this method; 23. After induction of Sequence 2 of this method; 24. Supernatant of Sequence 2 of this method; 25. Precipitate of Sequence 2 of this method; 26. Flow-through of nickel affinity column of Sequence 2 of this method; 27. Washing of impurities in nickel affinity column of Sequence 2 of this method; 28. Elution of nickel affinity column of Sequence 2 of this method.

[0027] Figure 5 SAS-PAGE results of a comparative expression experiment of human SMURF1 protein using codon sequences designed using this method and other codon sequences. 1. Optimized sequence using traditional algorithm - before induction; 2. Optimized sequence using traditional algorithm - after induction; 3. Optimized sequence using traditional algorithm - supernatant; 4. Optimized sequence using traditional algorithm - precipitation; 5. Optimized sequence using traditional algorithm - nickel affinity column flow-through; 6. Optimized sequence using traditional algorithm - nickel affinity column washing; 7. Optimized sequence using traditional algorithm - nickel affinity column elution; 8. Sequence designed using this method - before induction; 9. Sequence designed using this method - after induction; 10. Sequence designed using this method - supernatant; 11. Sequence designed using this method - precipitation; 12. Sequence designed using this method - nickel affinity column flow-through; 13. Sequence designed using this method - nickel affinity column washing; 14. Sequence designed using this method - nickel affinity column elution. Detailed Implementation

[0028] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, the present invention can be implemented in many different ways as defined and covered by the claims.

[0029] Example like Figure 1 As shown, the codon sequence design device based on a large-scale multimodal model includes the following four parts: (A) Multimodal information encoding module ( Figure 2 ) The method of this invention uses the amino acid sequence of a protein as the core input, and introduces its corresponding spatial folding and other auxiliary information to achieve multimodal fusion modeling, thereby providing rich contextual semantics and spatial feature support for high-quality codon sequence design.

[0030] Specifically, the method first extracts the primary sequence of the target protein, i.e., the linear arrangement of amino acid residues, as the basic input for the codon design task. This sequence directly determines the length and composition of the final generated codon sequence, constituting the "content axis" of the language model. At the same time, the model can also receive multimodal information actively provided by the user, or call the structure prediction module to automatically perform spatial conformation modeling on the input sequence and extract its high-dimensional spatial information.

[0031] The representation of protein structure information includes, but is not limited to: atomic coordinates of each amino acid residue, describing its precise position in three-dimensional space; inter-residue distance matrix, reflecting the spatial proximity of residues at different positions in the sequence; biophysical properties such as secondary structure type (e.g., α-helix, β-sheet, random coil) and solvent accessibility; and optional topological relationships or contact maps, used to capture contextual dependencies of distant but spatially adjacent sequences.

[0032] In model representation, amino acid sequences are encoded as standard token sequences and mapped to a high-dimensional vector space through an encoding layer. Multimodal information, based on different representations, is encoded as a set of auxiliary input vectors that can be processed by the Transformer, and input into the main model along with the sequence vectors. To achieve effective fusion of sequence-structure information, this invention supports multiple integration mechanisms, including parallel attention, cross-attention, and a fusion layer, thereby endowing the generative model with spatial awareness and contextual understanding capabilities.

[0033] The multimodal joint encoding strategy proposed in this invention can significantly enhance the ability of deep generative models to model and perceive the correlation between the local spatial conformation of target proteins and functionally relevant structural regions and codon sequences. When designing codons, the model can not only consider expression preferences but also better coordinate key factors such as sequence stability, translation efficiency, and structural fidelity.

[0034] (B) Generative module driven by large-scale multimodal models ( Figure 3 ) Building upon multimodal protein representation vectors, this invention designs and introduces a Large Multimodal Model (LMM) based on the Transformer architecture for structure-aware codon sequence design. This model learns codon usage patterns in natural proteins directly based on amino acid sequence context, folding, and other information, without relying on template genes or reference sequences. The model consists of three main parts: a context-aware amino acid-codon relationship modeling module, a protein folding association sensing module based on structural information fusion, and a temperature-sampling-based diverse sequence generation module, ensuring accurate, multimodal, and diverse codon design.

[0035] The model models the relationship between amino acid sequence context and codon selection through a context-aware generation mechanism. Employing a multi-layer self-attention mechanism, the model captures semantic dependencies between distant residues in the amino acid sequence, effectively learning the linguistic patterns of codon distribution and the contextual adaptability between synonymous codons, thus achieving a dual match between encoding efficiency and biological semantics. The model incorporates protein folding structure features (including residue spatial positions, contact relationships, secondary structures, etc.) as auxiliary inputs, enabling it to perceive the local spatial conformation of the target protein and adjust the generation strategy accordingly. This achieves structure-sensitive codon selection and enhances the structural fidelity and stability of the design results in protein expression. The model achieves diverse codon sequence designs through a temperature sampling mechanism. This invention introduces a temperature control parameter as a means of adjusting the generation probability distribution. By setting the temperature parameter, the model can output multiple equivalent candidate codon sequences while maintaining semantic correctness, thereby improving the diversity of the generated results and meeting the needs of subsequent personalized selection, fusion optimization, and experimental verification.

[0036] The codon sequence generation module is the core component of the overall method. Relying on the powerful knowledge induction and complex distribution modeling capabilities of large-scale language models, it breaks through the bottlenecks of traditional codon optimization methods in local fragment dependency modeling, spatial structure perception, and sequence diversity control, and achieves precise, multimodal, and diversified codon design.

[0037] The generation of codon sequences is achieved through a label loss function and model training, as detailed below: Label loss is the commonly used cross-entropy loss in Transformer architecture. Specifically, let the batch size be B, the total number of password subcategories be C, and the sample length be L; Represented as: in, It is the one-hot encoding of the target tag. This represents the model output.

[0038] The overall loss function in this embodiment is cross-entropy, but other losses may be added: The loss function can be adjusted based on the frequency or importance of different codons to prevent overfitting and improve the model's generalization ability. This loss function optimizes the quality and efficiency of protein codon design through weight adjustment and noise control, making it particularly suitable for cross-species expression optimization tasks.

[0039] During model training, the dataset was shuffled and randomly divided into training, validation, and test sets. The final model achieved an accuracy of 94.6% on the validation set and 91.8% on the test set.

[0040] (C) Sequence scoring and screening module To ensure the generated codon sequences possess good expression potential and biological rationality, this invention designs a multi-dimensional weighted sequence scoring and screening module for post-processing and quality control of the model output. This module introduces a series of widely used and biologically significant scoring indicators, including CAI (Conversion Ability Indicator), GC content control, and mRNA secondary structure stability prediction, to measure the mRNA quality of the generated codon sequences from multiple perspectives, as well as their expressibility and protein folding rationality in the target host system, comprehensively improving the practicality and reliability of the results. The final score is a weighted average of multiple evaluation indicators. By comprehensively scoring each candidate codon sequence, potentially inefficient expression or unstable design results can be effectively screened out, improving the overall output quality.

[0041] The module supports sorting and filtering mechanisms. This method scores and sorts candidate sequences based on the aforementioned multidimensional indicators, and outputs a set of high-scoring codon sequences for subsequent customized filtering, fusion optimization, or batch testing in industrial high-throughput expression experiments, thereby significantly improving the method's engineering adaptability and industrial practical value.

[0042] Experimental results 1. Recombinant expression experiment of human dihydrolactic acidase in Escherichia coli This method was used to design codons for human dihydrowheyase, and the designed codon sequences were compared with sequences optimized by traditional algorithms and the natural sequence. The traditional algorithm used for comparison was the GeneOptimizer codon optimization tool provided by Thermofisher. All three gene sequences were sent to GenScript for synthesis, and plasmid samples were returned one week later. The vector used was pET-28a+, and the antibiotic was kanamycin. The genes were transformed into *E. coli* (…). E. coli Competent cells prepared from BL21(DE3) strain were cultured in 250 ml of medium at 37°C for 4 h, then inoculated with IPTG for induction, and subsequently cultured at 16°C for 14-16 h. The OD value of the cultured bacterial suspension was measured, and an equal volume of the suspension was centrifuged, resuspended, and disrupted based on the OD value. Since a His tag was added during gene sequence synthesis, the collected supernatant protein was purified using a nickel affinity column. After passing the supernatant through the nickel column, it was washed sequentially with 20 ml of 20 mM imidazole solution and then eluted with 40 ml of 250 mM imidazole solution.

[0043] SDS-PAGE experiments were performed on samples before induction, after induction, supernatant, precipitate, flow-through, washing, and elution. The results are as follows: Figure 4 As shown in the figure, the SDS-PAGE experimental results show that the overall protein expression level of the codon sequence designed by this method is significantly higher than that of the optimized sequence and the original gene sequence of the traditional algorithm.

[0044] The protein content in each sample was further determined by ultraviolet absorption method, and the results are shown in Table 2.

[0045] Table 2. Results of repeated purification and quantification experiments of human dihydrolactic acidase The results showed that when the codon sequence designed using this method was used for recombinant expression of human dihydrolactic enzyme in *E. coli*, the protein yield was increased by 250% compared to the original sequence, and by 430% compared to the protein yield of codon sequence optimized by traditional algorithms. Protein quantification experiments fully validated that, compared to the natural sequence and traditional codon optimization methods, the codon sequence designed using this method can significantly improve the soluble expression level of human dihydrolactic enzyme.

[0046] 2. Recombinant expression experiment of human Smurf1 in Escherichia coli Smad ubiquitination regulator 1 (Smurf1) has attracted much attention in the field of biological research since its discovery. It can interact with key proteins in multiple signaling pathways, especially in the TGF-β signaling pathway, where it plays an important regulatory role in the ubiquitination and degradation of Smad proteins. This discovery has opened the door to the study of Smurf1 function.

[0047] The applicant used human Smurf1 as a research subject and attempted to solve the problem of soluble recombinant protein expression using a codon design / optimization strategy. Traditional methods failed to achieve detectable soluble expression of the target protein in *E. coli* with the Smurf1 codon sequence optimized for this purpose. However, the codon sequence designed using this method was successfully expressed in *E. coli*, and a large amount of soluble protein was detected. This result strongly demonstrates that this method can effectively improve the success rate of soluble expression of difficult-to-express proteins. Figure 5 ) 3. Conclusion This invention proposes a method and apparatus for deep codon sequence generation based on a large-scale multimodal model. It integrates protein amino acid sequence and spatial folding information to construct a multimodal representation. Through context awareness, structure regulation, temperature sampling, and confidence scoring, it achieves efficient and accurate codon design for target proteins. Compared to existing codon optimization methods, this invention has significant advantages in capturing amino acid context information, understanding the relationship between structure and codon selection, and outputting and screening diverse candidates, significantly improving the overall performance of the design results in terms of expression efficiency, stability, and diversity.

[0048] To verify the effectiveness of the method of this invention, the applicant first independently constructed a large-scale dataset of protein sequences, structures, and their genes, and then completed model training and testing based on this dataset. Experimental results show that the constructed model achieves a prediction accuracy of 94.6% on the validation set and 91.8% on the test set, demonstrating excellent generalization ability and stability.

[0049] Human dihydrolactic acid glycoside and human Smurf1 were further selected as test subjects. Recombinant expression experiments were conducted in *E. coli* BL21(DE3) competent cells using the natural sequence, the results of traditional codon optimization algorithms, and the codon sequence designed by this method, respectively. The results showed that the codon sequence designed by this method significantly improved protein expression levels and solubility: the final expression yield of human dihydrolactic acid glycoside was increased by 250% to 430% compared to the natural sequence, and soluble expression of Smurf1 in *E. coli*, which could not be expressed by traditional optimization methods, was achieved.

[0050] In summary, the structure-aware codon generation method proposed in this invention successfully generates highly expressive, stable, and adaptable codon sequences directly from protein information. This method addresses the urgent need of the biotechnology industry for intelligent gene design tools and is particularly suitable for several key application scenarios, including recombinant protein production, synthetic biology, nucleic acid drugs, nucleic acid vaccines, and transgenic engineering. Its efficiency and flexibility provide strong support for next-generation biodesign platforms, possessing broad promotional value and promising industrial application prospects.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A codon sequence design device based on a large-scale multimodal model, characterized in that, It includes the following three parts: (A) Multimodal information encoding module: including protein three-dimensional structure acquisition layer and amino acid sequence acquisition layer. The former is directly input into the protein three-dimensional structure encoding layer, or the amino acid sequence is input into the protein three-dimensional structure prediction layer when the experimental structure cannot be obtained, and the latter is directly input into the amino acid sequence encoding layer. (B) Generative module driven by a large multimodal model based on the Transformer architecture: including a codon sequence generator and a temperature sampling layer. The protein three-dimensional structure coding layer, amino acid sequence coding layer and sampling temperature are input into the codon sequence generator, which generates multiple candidate sequences through the temperature sampling layer. (C) Sequence scoring and screening module: The multiple candidate sequences generated by the generation module are scored using multidimensional weighted indicators, including various mRNA quality assessment indicators. High-scoring codon sequences are selected based on the scoring results.

2. A method for designing codon sequences based on a large-scale multimodal model using the apparatus described in claim 1, characterized in that, The specific steps are as follows: S1. Determine whether the protein's three-dimensional structure can be obtained from a public database based on the amino acid sequence: if the acquisition fails, obtain the protein's three-dimensional structure coding layer through the protein three-dimensional structure prediction layer; if the acquisition is successful, directly obtain the protein's three-dimensional structure coding layer from the database. S2. Input the protein three-dimensional structure coding layer, amino acid sequence coding layer, and sampling temperature into the codon sequence generator to generate codon sequences through a large-scale multimodal model; S3. Codon sequences are designed by using sequence scoring and screening based on multiple mRNA quality assessment indicators to obtain high-scoring codon sequences.

3. The method according to claim 2, characterized in that, In step S2, the generation of codon sequences is achieved through training a label loss function and a corresponding deep generative model.

4. The method according to claim 3, characterized in that, The label loss function is cross-entropy loss. And the sum of any possible additional loss functions.

5. The method according to claim 4, characterized in that, Let the batch size be B, the total number of password subcategories be C, and the sample length be L; Represented as: in, It is the one-hot encoding of the target tag. Indicates the model output; The overall loss function is the sum of all losses: 。 6. The method according to claim 3, characterized in that, During model training, the dataset is randomly divided into training, validation, and test sets.

7. The method according to claim 2, characterized in that, In step S3, the sequence scoring method is: sorting and filtering by multidimensional weighted indicators.

8. The method according to claim 3, characterized in that, In step S3, the sequence screening indicators include: CAI and GC content control.

Images