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

Abstract

The invention provides a method and a device for designing a protein coding codon sequence based on a large multi-modal model, which comprehensively utilize information such as an amino acid sequence, space folding and mRNA quality of protein to realize high-quality, controllable and evaluable codon sequence design. According to the method, the dependence of a traditional codon optimization strategy on a reference gene template is broken through, and a codon optimization problem is converted into a codon design task based on protein information. Through comprehensive consideration of information such as space folding of protein, context of amino acid sequences, mRNA quality and the like, the model can generate a codon sequence which is more in line with codon use preference of specific species, reasonable in structure and efficient in expression, so that a new thought and technical support are provided for safe, stable, efficient and controllable protein expression.

Description

Technical Field

[0001] The present invention relates to the technical field of protein coding codon design, and in particular, to a codon sequence design method and device based on a large multimodal model. Background Art

[0002] Proteins play crucial roles as both participants and executors in life, participating extensively in diverse processes such as cellular structure, metabolism, signal transduction, and immune response. Controlled protein production holds broad value in diverse applications, including recombinant protein drug development, vaccine design, gene therapy, and synthetic biology research. It is also crucial for the development of cutting-edge biotechnologies and the promotion of related industries.

[0003] The process of protein production in cells is essentially the synthesis of peptide chains by ribosomes based on the genetic information carried by messenger RNA (mRNA). In biology, a codon refers to the three consecutive bases in mRNA or the corresponding deoxyribonucleic acid (DNA) sequence that specify an amino acid. During the translation process from gene to mRNA and then to protein, each codon encodes a specific amino acid. Nature contains four types of bases: adenine (A), cytosine (C), guanine (G), and thymine (T). These three bases can be combined to form 64 different codons, corresponding to only 20 common amino acids, plus a signal to terminate peptide production, for a total of 21 types of information. Therefore, there is redundancy in the codon encoding of amino acids. With the exception of methionine and tryptophan, the remaining 18 essential amino acids are encoded by two to six codons. Different codons encoding the same amino acid are synonymous.

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

[0005] Some of the currently known regularity factors include: 1) Different species have different preferences for codon selection. The corresponding codon optimization method is to improve 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 will cause mRNA to form a complex secondary structure, and too low 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 possibility of obtaining correctly folded proteins.

[0006] However, some of the above principles are too general and not operational, 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, such methods often use the codon usage frequency of the host cell as a primary reference, expecting high-frequency codons to lead to high-speed peptide chain synthesis, but ignoring the context of the gene sequence; second, studies have shown that current mainstream codon optimization strategies may interfere with the folding conformation and stability of proteins, thereby changing 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 (AI) technology, attempts have been made to apply deep learning methods to the field of codon optimization. Specifically, neural network models are constructed to learn the mapping between gene sequence and expression levels, thereby achieving intelligent optimization of codon sequences. Representative work includes 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 of proteins, thus failing to address the disconnect between peptide chain expression and spatial folding. Furthermore, due to the limited availability of specific experimental data on protein expression, the extremely complex data sources, and the significant differences in methodological systems, it is difficult to train truly effective deep learning models. Summary of the Invention

[0008] The present invention aims to provide a codon sequence design method and apparatus based on a large multimodal model to address the limitations of traditional codon optimization algorithms.

[0009] To achieve the above objectives, the present invention provides a codon sequence design device based on a large multimodal model, comprising the following four parts: (A) Multimodal information encoding module: This module includes a protein 3D structure acquisition layer and an amino acid sequence acquisition layer. The former is directly input into the protein 3D structure encoding layer, or the amino acid sequence is input into the protein 3D structure prediction layer when the experimental structure is unavailable. The latter is directly input into the amino acid sequence encoding layer. (B) Large-scale multimodal model-driven generation module: This module includes a codon sequence generator and a temperature sampling layer. The protein three-dimensional structure encoding layer, amino acid sequence encoding layer, and sampling temperature are input into the codon sequence generator, which then generates multiple candidate sequences through the temperature sampling layer. (C) Sequence scoring and screening module: Multiple candidate sequences generated by the generation module are scored using multidimensional weighted indicators, including multiple mRNA quality assessment indicators, and high-scoring designed codon sequences are screened based on the scoring results.

[0010] The present invention also provides a codon sequence design method based on a large multimodal model implemented using the aforementioned device, the specific steps of which are as follows: S1. Determine whether the protein's 3D structure can be obtained from a public database based on the amino acid sequence. If the 3D structure cannot be obtained, use protein 3D structure prediction to obtain the protein's 3D structure encoding. If the 3D structure can be obtained, obtain the protein's 3D structure encoding directly from the database. S2. Input the protein 3D structure code, amino acid sequence code, and sampling temperature into the codon sequence generator, and generate the codon sequence using a large multimodal model; S3. Codon sequences were scored and screened based on multiple mRNA quality assessment metrics to obtain high-scoring designed codon sequences.

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

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

[0013] As one of the further preferred technical solutions, let the batch size be B, the total number of codon categories be C, and the sample length be L; Expressed as: in, is the one-hot encoding of the target label, Represents the model output.

[0014] The total loss function is the sum of all losses: As one of the further preferred technical solutions, the data set is randomly divided into a training set, a validation set and a 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 multi-dimensional weighted indicators.

[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: The present invention provides a codon sequence design method and device based on a large multimodal model, which comprehensively utilizes information such as the amino acid sequence, spatial folding, and mRNA quality of the protein to achieve high-quality, controllable, and evaluable codon sequence design. The present invention breaks through the reliance of traditional codon optimization strategies on reference gene templates and transforms the codon optimization problem into a codon design task based on protein information. By comprehensively considering information such as the spatial folding structure of the protein, the context of the amino acid sequence, and the quality of the mRNA, the model can generate codon sequences that are more consistent with the codon usage preferences of specific species, have a reasonable structure, and are highly efficient in expression, thereby providing new ideas and technical support for safe, stable, efficient, and controllable protein expression.

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

[0019] First, the present invention designs the codon sequence directly from the amino acid sequence without the need for a template codon sequence, achieving end-to-end precise design and flexible adaptation, and getting rid of the dependence on the starting reference template codon sequence.

[0020] Secondly, the present invention comprehensively considers information such as the amino acid sequence context of the protein and the spatial folding of the protein, which can enable the model to better perceive the relationship between codon usage and various influencing factors, and improve the expression success rate of the generated codon sequence.

[0021] Finally, the model of the present 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 above-described objects, features and advantages, the present invention has other objects, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings: Figure 1 Schematic diagram of the deep codon sequence generation method based on large multimodal models.

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

[0025] Figure 3 Schematic diagram of the codon sequence generation module based on a large multimodal model.

[0026] Figure 4SAS-PAGE results of the expression comparison experiment of the codon sequence designed by this method and other codon sequences for human dihydroorotase. 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 wash; 7. Original sequence - nickel affinity column elution; 8. Sequence optimized by traditional algorithm - before induction; 9. Sequence optimized by traditional algorithm - after induction; 10. Sequence optimized by traditional algorithm - supernatant; 11. Sequence optimized by traditional algorithm - precipitate; 12. Sequence optimized by traditional algorithm - nickel affinity column flow-through; 13. Sequence optimized by traditional algorithm - nickel affinity column wash; 14. Sequence optimized by traditional algorithm - nickel affinity column elution; 15. Sequence designed by this method 1 - before induction; 16 .The method is designed to have sequence 1 - after induction; 17. The method is designed to have sequence 1 - supernatant; 8. The method is designed to have sequence 1 - precipitation; 19. The method is designed to have sequence 1 - nickel affinity column flow-through; 20. The method is designed to have sequence 1 - nickel affinity column wash; 21. The method is designed to have sequence 1 - nickel affinity column elution; 22. The method is designed to have sequence 2 - before induction; 23. The method is designed to have sequence 2 - after induction; 24. The method is designed to have sequence 2 - supernatant; 25. The method is designed to have sequence 2 - precipitation; 26. The method is designed to have sequence 2 - nickel affinity column flow-through; 27. The method is designed to have sequence 2 - nickel affinity column wash; 28. The method is designed to have sequence 2 - nickel affinity column elution.

[0027] Figure 5 SAS-PAGE results of a comparative expression experiment comparing the codon sequences designed using this method with other codon sequences for human SMURF1 protein. 1. Sequence optimized using the traditional algorithm - before induction; 2. Sequence optimized using the traditional algorithm - after induction; 3. Sequence optimized using the traditional algorithm - supernatant; 4. Sequence optimized using the traditional algorithm - precipitate; 5. Sequence optimized using the traditional algorithm - flow-through from the nickel affinity column; 6. Sequence optimized using the traditional algorithm - wash from the nickel affinity column; 7. Sequence optimized using the traditional algorithm - elution from the nickel affinity column; 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 - precipitate; 12. Sequence designed using this method - flow-through from the nickel affinity column; 13. Sequence designed using this method - wash from the nickel affinity column; 14. Sequence designed using this method - elution from the nickel affinity column. DETAILED DESCRIPTION

[0028] The embodiments of the present invention are 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 in FIG, the codon sequence design device based on the large multimodal model includes the following four parts: (A) Multimodal information encoding module ( Figure 2 ) The method of the present invention takes the amino acid sequence of the protein as the core input, and introduces its corresponding spatial folding as 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—the linear order of amino acid residues—as the fundamental input for the codon design task. This sequence directly determines the length and composition of the resulting codon sequence, forming the "content axis" of the language model. Simultaneously, the model can also accept multimodal information proactively provided by the user or invoke a structure prediction module to automatically perform spatial conformational modeling on the input sequence, extracting its high-dimensional spatial information.

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

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

[0033] The multimodal co-encoding strategy proposed in this paper significantly enhances the deep generative model's ability to model and perceive the local spatial conformation of the target protein and the correlation between functionally relevant structural regions and codon sequences. This model not only considers expression preferences when designing codons, but also better balances key factors such as sequence stability, translation efficiency, and structural fidelity.

[0034] (B) Large multimodal model-driven generation module ( Figure 3 ) Based on multimodal protein representation vectors, this paper designs and introduces a Transformer-based large multimodal model (LMM) for structure-aware codon sequence design. This model learns codon usage patterns in natural proteins directly from amino acid sequence context, folding, and other information, without relying on template genes or reference sequences. The model consists of three main components: a context-aware amino acid-codon relationship modeling module, a protein folding association perception module that integrates structural information, and a temperature-sampling diverse sequence generation module, ensuring precise, multimodal, and diverse codon design.

[0035] The model models the relationship between amino acid sequence context and codon selection through a context-aware generative 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, thereby achieving a dual match between coding efficiency and biological semantics. By introducing protein folding structural features (including residue spatial positions, contact relationships, secondary structures, etc.) as auxiliary input, the model can perceive the local spatial conformation of the target protein and adjust the generation strategy accordingly, achieving structure-sensitive codon selection and enhancing the structural fidelity and stability of the design results in protein expression; The model achieves diverse codon sequence design through a temperature sampling mechanism. This paper introduces a temperature control parameter as a means of regulating the generated probability distribution. By setting the temperature parameter, the model can output multiple equivalent candidate codon sequences while maintaining semantic correctness, thereby increasing the diversity of generated results and meeting the needs of subsequent personalized screening, fusion optimization, and experimental verification.

[0036] This 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 realizes the precision, multimodality, and diversification of codon design.

[0037] The generation of codon sequences is achieved through label loss function and model training, as follows: Label loss is the cross entropy (CE, CrossEntropy) loss commonly used in Transformer architecture Specifically, let the batch size be B, the total number of codon categories be C, and the sample length be L; Expressed as: in, is the one-hot encoding of the target label, Represents the model output.

[0038] The total loss function in this embodiment is cross entropy, but it does not exclude the addition of other losses: The loss function can be adjusted based on the frequency or importance of different codons to prevent overfitting and improve the generalization ability of the model. 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 screening module In order to ensure that the generated codon sequence has good expression potential and biological rationality, the present invention has designed a set of multi-dimensional weighted sequence scoring screening modules for post-processing and quality control of the model output results. This module introduces a series of widely used and biologically meaningful scoring indicators, including CAI, GC content control, mRNA secondary structure stability prediction, etc., which measure the mRNA quality of the generated codon sequence, as well as the expressibility and protein folding rationality in the target host system from multiple angles, and comprehensively improve 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, it is possible to effectively screen out design results with potential inefficient expression or insufficient stability, thereby improving the overall output quality.

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

[0042] Experimental results 1. Recombinant expression of human dihydroorotase in Escherichia coli This method was used to design the coding codons for human dihydroorotase, and the designed codon sequences were compared with the 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 resistance was kanamycin. The gene was transformed into Escherichia coli ( E. coli Competent cells prepared from the BL21(DE3) strain were cultured in 250 ml of medium at 37°C for 4 hours, then induced with IPTG. The cells were then transferred to 16°C for 14-16 hours. The OD value of the cultured cells was measured, and equal amounts of the culture were centrifuged, resuspended, and disrupted based on the OD value. Because a His tag was added during the synthesis of the gene sequence, the collected supernatant protein was purified using a nickel affinity column. After passing the supernatant through the nickel column, it was washed 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 and after induction, supernatant, precipitate, flow-through, wash and elution. The experimental results are shown in Figure 2. Figure 4 The SDS-PAGE experimental results show that the overall protein expression of the codon sequence designed by this method is significantly higher than that of the traditional algorithm optimized sequence and the original gene sequence.

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

[0045] Table 2. Results of repeated purification and quantitative experiments of human dihydroorotase Results showed that the codon sequence designed using this method for recombinant expression of human dihydroorotase in E. coli increased protein yield by 250% compared to the native sequence and by 430% compared to the codon sequence optimized using a traditional algorithm. Protein quantification experiments fully demonstrated that the codon-designed sequence based on this method significantly increased the soluble expression of human dihydroorotase compared to the native sequence and traditional codon optimization methods.

[0046] 2. Recombinant expression of human Smurf1 in Escherichia coli Since its discovery, Smad ubiquitination regulator 1 (Smurf1) has garnered significant attention in biological research. It interacts with key proteins in multiple signaling pathways, particularly in the TGF-β signaling pathway, where it plays a crucial role in regulating the ubiquitination and degradation of Smad proteins. This discovery opened the door to further research on Smurf1's function.

[0047] The applicant used human Smurf1 as a research target and attempted to employ a codon design / optimization strategy to address the issue of soluble recombinant protein expression. The Smurf1 codon sequence optimized using traditional methods was completely unable to achieve detectable soluble expression of the target protein in E. coli. However, the codon sequence designed using this method was successfully expressed in E. coli, and a significant 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 paper proposes a method and device for deep codon sequence generation based on a large-scale multimodal model. This method integrates protein amino acid sequences with spatial folding information to construct a multimodal representation. Through contextual awareness, structural control, temperature sampling, and confidence scoring screening, this method enables efficient and accurate codon design for target proteins. Compared to existing codon optimization methods, this method offers significant advantages in capturing amino acid contextual information, understanding the relationship between structure and codon selection, and outputting and screening diverse candidates. This significantly improves the overall performance of the design results in terms of expression efficiency, stability, and diversity.

[0048] To validate the effectiveness of the method, the applicants first independently constructed a large-scale dataset of protein sequences, structures, and genes, and then trained and tested the model based on this dataset. The experimental results showed that the constructed model achieved a prediction accuracy of 94.6% on the validation set and 91.8% on the test set, demonstrating excellent generalization and stability.

[0049] Human dihydroorotase and human SMURF1 were further tested, and recombinant expression experiments were conducted in E. coli BL21 (DE3) competent cells using the native sequence, the codon sequences obtained by traditional codon optimization algorithms, and the codon sequences designed by this method. The results showed that the codon sequences designed by this method significantly improved protein expression and solubility: the final expression yield of human dihydroorotase increased by 250% to 430% compared to the native sequence, and Smurf1, which was difficult to express using traditional optimization methods, was expressed soluble in E. coli.

[0050] In summary, the structure-aware codon generation method proposed in this paper successfully generates high-expression, high-stability, and high-adaptability codon sequences directly from protein information. This method responds to the biotechnology industry's urgent need for intelligent gene design tools and is particularly suitable for multiple key application scenarios, including recombinant protein production, synthetic biology, nucleic acid drugs, nucleic acid vaccines, and transgenic engineering. Its high efficiency and flexibility provide strong support for next-generation biodesign platforms and have broad application potential.

[0051] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A codon sequence design device based on a large multimodal model, characterized in that: It includes the following four parts: (A) Multimodal information encoding module: This module includes a protein 3D structure acquisition layer and an amino acid sequence acquisition layer. The former is directly input into the protein 3D structure encoding layer, or the amino acid sequence is input into the protein 3D structure prediction layer when the experimental structure is unavailable. The latter is directly input into the amino acid sequence encoding layer. (B) Large-scale multimodal model-driven generation module: This module includes a codon sequence generator and a temperature sampling layer. The protein three-dimensional structure encoding layer, amino acid sequence encoding layer, and sampling temperature are input into the codon sequence generator, which then generates multiple candidate sequences through the temperature sampling layer. (C) Sequence scoring and screening module: Multiple candidate sequences generated by the generation module are scored using multidimensional weighted indicators, including multiple mRNA quality assessment indicators, and high-scoring designed codon sequences are screened based on the scoring results.

2. A method for designing codon sequences based on a large multimodal model implemented using the apparatus of claim 1, characterized in that: The specific steps are as follows: S1. Determine whether the protein's 3D structure can be obtained from a public database based on the amino acid sequence. If the 3D structure cannot be obtained, use protein 3D structure prediction to obtain the protein's 3D structure encoding. If the 3D structure can be obtained, obtain the protein's 3D structure encoding directly from the database. S2. Input the protein 3D structure code, amino acid sequence code, and sampling temperature into the codon sequence generator, and generate the codon sequence using a large multimodal model; S3. Codon sequences were scored and screened based on multiple mRNA quality assessment metrics to obtain high-scoring designed codon sequences.

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

4. The method according to claim 3, characterized in that The label loss function is the 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 codon categories be C, and the sample length be L; Expressed as: in, is the one-hot encoding of the target label, Represents the model output; The total loss function is the sum of all losses: 。 6. The method according to claim 3, characterized in that When training the model, the dataset is randomly divided into training set, validation set and test set.

7. The method according to claim 2, characterized in that In step S3, the sequence scoring method is: sorting and screening of multi-dimensional weighted indicators.

8. The method according to claim 3, characterized in that In step S3, sequence screening indicators include but are not limited to: CAI, GC content control, etc.

Images