A gene synonymous codon optimization method, system, computer device and medium

Abstract

The application provides a gene synonymous codon optimization method, system, computer device and medium, and belongs to the field of bioinformatics and synthetic biology, and the method comprises the following steps: a first model takes an original gene sequence as input, outputs monomer ribosome density distribution of each codon site, and quantifies background translation efficiency; a second model takes a candidate sequence and the calibrated first model output as input in an optimization process, outputs dimer ribosome density distribution through feature fusion, and serves as a scoring basis to accurately locate ribosome collision risk; and the first model takes the optimized sequence as input, independently outputs dimer density distribution, and performs orthogonal verification. Through the construction of two deep learning models sharing a backbone network, the application realizes translation density prediction and targeted optimization of a full-length gene sequence, realizes global optimization of the whole gene and result reliability verification under the premise that the amino acid sequence remains unchanged, and provides an efficient sequence design tool for genetic engineering and synthetic biology.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and synthetic biology, and specifically relates to a method, system, computer equipment, and medium for optimizing gene synonymous codons. Background Technology

[0002] Synonymous codon optimization reveals the regulatory rules of the translation process by analyzing the molecular causes of ribosome collisions and the quality control mechanism of cellular translation. It provides efficient gene optimization strategies for biopharmaceuticals (such as insulin and monoclonal antibodies) and synthetic biology (such as biofuels and industrial enzyme preparations). By improving the expression yield and translation efficiency of target proteins, it significantly optimizes the production efficiency of microbial cell factories, thereby reducing the research and development and production costs of bioproducts.

[0003] Existing methods generally employ Transformer or convolutional models for local encoding and local optimization strategies, resulting in Ribo-seq density prediction. This involves predicting the translation density score of the central site based on a short window sequence, and suppressing peaks only in the target region during optimization. This lacks global modeling and overall inspection of the entire gene or complete CDS. Furthermore, the model cannot clearly show the source of signal changes during optimization, making it difficult to distinguish whether density fluctuations are caused by changes in background translation efficiency or by actual ribosome collision enhancement.

[0004] The aforementioned technical deficiencies make it difficult to directly apply the optimization results to the design and modification of full-length genes. Furthermore, the ambiguity of the signal source leads to poor interpretability and low reliability of the optimization process, ultimately limiting the reliable application of existing methods in practical genetic engineering and synthetic biology. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method, system, computer device, and medium for optimizing gene synonymous codons.

[0006] To achieve the above objectives, the present invention provides a method for optimizing gene synonymous codons, comprising: Collect the codon index sequence of the original gene of the target organism.

[0007] The codon index sequence is input into the first backbone network RiboMamba, and the monosomal ribosome density distribution of the codon index sequence is output. The first backbone network RiboMamba consists of an input layer, an embedding module, a positional encoding, a multi-scale convolutional block, a Mamba state space block, and an output layer connected in sequence.

[0008] Based on the peak position or preset target region in the single ribosome density distribution, an optimal window is determined. A codon site is randomly selected within the optimal window for synonymous codon replacement, resulting in an updated codon index sequence. The updated codon index sequence is input into the second backbone network RiboMamba, outputting a first dual-ribosome density distribution. The second backbone network RiboMamba includes: a fusion module added between the multi-scale convolutional blocks and Mamba state space blocks of the first backbone network RiboMamba; a score is calculated based on the first dual-ribosome density distribution, and when the score meets preset conditions, the current optimal codon index sequence is output; the second dual-ribosome density distribution of the current optimal codon index sequence is extracted again using the first backbone network RiboMamba.

[0009] If the density distribution of the second dimer ribosome matches the density distribution of the first dimer ribosome, then the current optimal codon index sequence is confirmed as the final optimal codon index sequence.

[0010] Preferably, the codon index sequence is input into the first backbone network RiboMamba, and the monomer ribosome density distribution of the codon index sequence is output. Specifically, this includes: mapping each codon index in the codon index sequence to a multi-dimensional encoding vector using multi-scale convolutional blocks, and adding positional encoding information to the encoding vector of each dimension to obtain an initial encoding sequence representation; capturing and mapping the long-range dependency encoding information of the initial encoding sequence representation using Mamba state space blocks to obtain the monomer ribosome density distribution of the codon index sequence.

[0011] Preferably, the fusion module includes: a splicing layer, a fully connected layer, and a GELU activation function; the initial encoding sequence representation of the updated codon index sequence is extracted using the second backbone network RiboMamba; the fusion module of the second backbone network RiboMamba splices the single ribosome density distribution and the initial encoding sequence representation to obtain fused encoding features; and the first dual ribosome density distribution is output based on the fused encoding features.

[0012] Preferably, the first backbone network RiboMamba is trained using two types of data samples: one is trained independently using monomeric ribosome density data to predict and output the monomeric ribosome density distribution; the other is trained independently using disosome ribosome density data to predict and output the second disosome ribosome density distribution; the monomeric ribosome density data and disosome ribosome density data are extracted from the codon index sequence of the original gene of the target organism.

[0013] Preferably, the scoring based on the density distribution of the first dimer ribosome specifically includes: using the sum of the density values ​​of the first dimer ribosomes at all codon sites within the optimal window as the score.

[0014] Preferably, determining the optimal window based on the peak position in the monomer ribosome density distribution or a preset target region further includes setting an out-of-window penalty term to suppress the generation of new peaks outside the optimal window.

[0015] Preferably, before splicing the monomer ribosome density distribution and the initial coding sequence representation, the method further includes: performing global linear calibration on the monomer ribosome density distribution output by the first backbone network RiboMamba, and splicing the calibrated monomer ribosome density distribution and the initial coding sequence representation.

[0016] This invention also provides a gene synonym codon optimization system, comprising: The data acquisition module is used to collect the codon index sequence of the original genes of the target organism.

[0017] A feature extraction module is used to input the codon index sequence into a first backbone network, RiboMamba, and output the single-unit ribosome density distribution of the codon index sequence. The first backbone network, RiboMamba, consists of an input layer, an embedding module, a positional encoding, a multi-scale convolutional block, a Mamba state space block, and an output layer connected in sequence. Based on the peak position or a preset target region in the single-unit ribosome density distribution, an optimal window is determined, and a codon site is randomly selected within the optimal window for synonymous codon replacement, resulting in an updated codon index sequence. The updated codon index sequence is then input into a second backbone network, RiboMamba, and outputs a first two-unit ribosome density distribution. The second backbone network, RiboMamba, includes a fusion module added between the multi-scale convolutional block and the Mamba state space block of the first backbone network, RiboMamba. The first two-unit ribosome density distribution is scored, and when the score meets a preset condition, the current optimal codon index sequence is output. The second two-unit ribosome density distribution of the current optimal codon index sequence is extracted again using the first backbone network, RiboMamba.

[0018] An optimization module is used to confirm the current optimal codon index sequence as the final optimal codon index sequence if the density distribution of the second dimer ribosome matches the density distribution of the first dimer ribosome.

[0019] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement any of the steps in the gene synonym codon optimization method.

[0020] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute any of the steps in the gene synonym codon optimization method.

[0021] The gene synonym codon optimization method provided by this invention has the following beneficial effects: This invention first collects the codon index sequence of the original gene of the target organism using the first backbone network, RiboMamba, and performs multidimensional and positional encoding on it. It then extracts the local contextual features and long-range dependency information of the entire gene, outputting the monomer ribosome density distribution at each codon site. This achieves global modeling of the full-length gene and quantification of background translation efficiency, overcoming the limitations of existing methods that can only perform local predictions based on short windows and lack overall inspection. Secondly, after determining the optimization window, candidate sequences are generated through synonymous codon substitution based on the second backbone network, RiboMamba. The encoded representation of the updated sequence is then concatenated and fused with the pre-acquired monomer density distribution. Based on the fusion features, the dimer ribosome density distribution is output and used as a scoring criterion. This clearly distinguishes between changes in background translation efficiency and actual ribosome collision signals during the optimization process, solving the problems of ambiguous signal sources and uninterpretable optimization processes. Finally, the optimal sequence is iteratively searched based on the scoring results and judged using the second dimer ribosome density distribution predicted by the first backbone network, RiboMamba, ensuring more accurate optimization results that can be directly applied to the design and modification of full-length genes. Attached Figure Description

[0022] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. 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.

[0023] Figure 1 This is a flowchart of a gene synonym codon optimization method according to an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0025] This invention provides a method for optimizing gene synonymous codons, specifically as follows: Figure 1 As shown, it includes the following steps: S1. Collect the codon index sequence of the original gene of the target organism.

[0026] First, the CDS sequence is encoded into a full-length codon index sequence, and a gene data structure is constructed. Then, the P-site offset is determined according to the specific experimental setup to ensure correct alignment of the ribosome signal with the sequence position. During the training and inference phases, the density values ​​are first non-negatively pruned and then transformed using a log1p transform to improve numerical stability and reduce the impact of extreme values ​​on model learning. The logarithmic smoothing transformation formula for the original ribosome count data is as follows:

[0027] ; Where i represents the codon position index in the full-length sequence of the target gene, and is a positive integer; This represents the count of raw ribosomes aligned to the i-th codon position in the sequencing data, and is a non-negative integer; 1 is a smoothing constant used to prevent raw counts from being detected. When = 0, the logarithmic function is undefined; This represents a logarithmic function with the natural constant e as its base. This represents the relative ribosome density value at the i-th codon position after logarithmic smoothing transformation.

[0028] S2. Input the codon index sequence into the first backbone network RiboMamba, and output the single-unit ribosome density distribution of the codon index sequence. The first backbone network RiboMamba consists of an input layer, an embedding module, a positional encoder, a multi-scale convolutional block, a Mamba state space block, and an output layer connected in sequence. Based on the peak position or a preset target region in the single-unit ribosome density distribution, determine the optimal window, and cyclically select a codon site within the optimal window for synonymous codon replacement to obtain the updated codon index sequence. Input the updated codon index sequence into the second backbone network RiboMamba, and output the first two-unit ribosome density distribution. The second backbone network RiboMamba includes: a fusion module added between the multi-scale convolutional block and the Mamba state space block of the first backbone network RiboMamba. Score the first two-unit ribosome density distribution, and when the score meets the preset conditions, output the current optimal codon index sequence. Use the first backbone network RiboMamba to extract the second two-unit ribosome density distribution of the current optimal codon index sequence again.

[0029] In this invention, the first and second backbone networks, RiboMamba and RiboMamba respectively, share the same custom network. This network is built from publicly available deep learning basic modules, and this invention names the custom network formed by combining these modules RiboMamba. The input of RiboMamba is the codon index sequence corresponding to the full-length CDS, and the sequence length is equal to the number of codons in the gene; the output of RiboMamba is a continuous value sequence corresponding to the input length, outputting a predicted density value (monomer ribosome density distribution) for each codon position, thereby obtaining the density distribution of the entire gene.

[0030] However, the first backbone network RiboMamba of this invention includes two training modes, each corresponding to different application scenarios: The first training mode: Monomer ribosome density data is extracted from the codon index sequence of the original gene of the target organism as a supervision label for training, and the first network instance is obtained. This instance is used to extract the monomer density distribution from the original gene sequence, characterize the background congestion during the translation process, and provide a reference for subsequent optimization.

[0031] The second training mode: The dimer ribosome density data is extracted from the codon index sequence of the original gene of the target organism as a supervision label for training, and a second network instance is obtained. The training process of this instance is independent of the first instance and does not rely on any background density information. It predicts the dimer density distribution based only on the sequence itself, which is used to orthogonally verify the optimization results and ensure the independence and objectivity of the verification process.

[0032] For the first and second backbone networks of RiboMamba with purely sequential input, the actual application data flow is closely integrated with the network structure, as detailed below: The first backbone network, RiboMamba, takes the codon index sequence of the original gene of the target organism as input. It maps each codon index to a 64-dimensional vector through an embedding layer and superimposes positional encoding to inject positional information, with the output length consistent with the input. Next, it enters a multi-scale convolutional feature block containing three parallel one-dimensional convolutional branches (kernel sizes of 3, 7, and 15, respectively). Each branch uses two layers of one-dimensional convolution combined with a normalization layer, GELU activation, and dropout. The three outputs are concatenated in the channel dimension and weighted by channel attention. Then, they are projected back to 64 dimensions through a linear layer and summed with the residuals of the convolutional block input to enhance stability. The above features are then input into a 6-layer Mamba state space block for full-length modeling. Each layer uses a pre-normalized structure, including Mamba state space operators, dropout, and residual connections. Finally, a positional regression output head (with a normalized layer followed by two linear mapping layers, using GELU activation and dropout in between) maps the features at each position to a continuous value, outputting the monomer ribosome density distribution at each codon site. This density distribution, as a quantification of background translation efficiency, is stored and used in subsequent optimization steps.

[0033] The second backbone network, RiboMamba, is used for collision prediction during the optimization process. In each iteration, a codon site is randomly selected within the optimization window for synonymous codon replacement, resulting in an updated candidate sequence. This candidate sequence is then input into the second backbone network, whose structure is essentially the same as the first backbone network, but a fusion module is added between the multi-scale convolutional blocks and the Mamba blocks. The candidate sequence passes through embedding layers, positional encoding, and multi-scale convolutional blocks sequentially to obtain sequence features. Simultaneously, the previously stored single-unit density distribution output from the first backbone network is globally linearly calibrated and then upscaled to 64 dimensions through a linear mapping layer to obtain density features. The sequence features and density features are concatenated along the channel dimension, projected back to 64 dimensions through a linear layer, and GELU activation is applied to form fused features. The fused features are then input into a 6-layer Mamba state space block for long-range modeling, and finally, the first dual-body ribosome density distribution for each codon site is output by the output head. A score within the optimization window (e.g., the sum of density values ​​within the window) is calculated based on this distribution, serving as the criterion for judging the quality of this modification. The optimization process is iterated until the score meets a preset threshold.

[0034] S3. If the density distribution of the second dimer ribosome matches the density distribution of the first dimer ribosome, then the current optimal codon index sequence is confirmed as the final optimal codon index sequence.

[0035] After optimization, the optimal candidate sequence is input again into the first backbone network RiboMamba (a second network instance) trained independently with dual-body density data. This instance does not rely on any background information and outputs the second dual-body ribosome density distribution based solely on the sequence itself. If the second dual-body density distribution matches the first dual-body density distribution output by the second backbone network during optimization (e.g., the integral difference within the window is less than a threshold or the correlation coefficient is greater than a threshold), the optimization result is confirmed to be valid, and the final sequence and mutation list are output.

[0036] Regarding specific model differences, the second backbone network, RiboMamba, is a fused input model. Following the aforementioned multi-scale convolutional blocks and before entering the Mamba state space block, the second backbone network, RiboMamba, introduces a density flow branch based on early fusion through a fusion module. The fusion module includes a concatenation layer, a fully connected layer, and a GELU activation function. It converts the output of the first backbone network, RiboMamba, after global linear calibration, into a 64-dimensional representation through a positional linear mapping, and concatenates it with the full-length features output from the sequence branch after convolution along the channel dimension. The concatenated features are projected back to 64 dimensions through a fully connected layer, and after a non-linear transformation by the GELU activation function, are fed into the subsequent Mamba state space block. This allows the Mamba module to simultaneously reference background density information and sequence context when modeling long sequences.

[0037] During the calibration phase, a batch of gene samples is first used to fit a linear calibration relationship so that the value of the Monosome true density after log1p transformation satisfies an approximately linear relationship with the prediction result of the first backbone network RiboMamba after secondary training. After obtaining the global calibration coefficients, the output of the first backbone network RiboMamba is first calibrated in the subsequent optimization and scoring phase, and then the calibrated result is used as the input of the second backbone network RiboMamba.

[0038] The optimization process takes the target gene, optimization window range, RiboMamba (first backbone network) / RiboMamba (second backbone network) models, calibration parameters, and synonymous codon mapping table as inputs. Under constraints, only synonymous codon substitutions that do not alter the amino acid sequence are allowed, and the maximum number of mutations or prohibited substitution regions can be further limited. In each iteration, the algorithm first randomly selects codon sites from the allowed positions, then samples substitutions from the corresponding synonym set to generate candidate sequences. Subsequently, the algorithm uses the output of the calibrated first backbone network RiboMamba as the input of the second backbone network RiboMamba to predict the Disome density distribution of the candidate sequences and calculate the objective function value to obtain a score. Finally, it searches for a better solution.

[0039] The sum of the first dimer ribosome densities at all codon sites within the optimization window is calculated as the score. The objective function formula for the optimization window score is as follows:

[0040] in, This represents the total score of the current candidate sequence within the optimization window, characterizing the overall congestion level of binocular ribosomes in this region; and These represent the start and end indexes of the optimization window within the full-length codon sequence, respectively; j represents the relative codon position index within the optimization window. This represents the first disosome ribosome density value at the j-th codon site of the current candidate sequence predicted by the second backbone network RiboMamba.

[0041] The sum of the ribosome density values ​​at all codon sites outside the optimization window is calculated, multiplied by a penalty coefficient, and added to the total score. The penalty coefficient is set to 0.5. A higher sum of density values ​​outside the window results in a higher total score and a lower probability of acceptance for the candidate sequence, thus suppressing the generation of new peaks outside the window. If the score is lower than the current score or meets the preset acceptance criteria, the current candidate sequence is updated to a better solution, and the iteration continues until the score falls below a set threshold or the maximum number of iterations is reached. The formula for the preset acceptance criteria is:

[0042] ; in, This represents the probability of accepting a candidate sequence with a worse score as the current better solution, and its value ranges from (0,1). This represents an exponential function with the natural constant e as its base. This represents the score of the candidate sequence after synonymous codon substitution. This indicates the current sequence score before the mutation; This represents the system parameters at the k-th iteration, which gradually approach 0 with increasing iteration number according to a preset annealing rate; when the generated random number interval [0, 1) is less than If the preset acceptance condition is met, then the threshold can be dynamically set based on the initial score of the target gene and the optimization target, for example, set to 80% of the initial score or a fixed value preset based on experience.

[0043] In the results validation phase, a Guardrail global health check is first performed. For candidate schemes generated from multiple genes and multiple random seeds, indicators such as the peak value change within the target window, the peak value change outside the window, and the global distribution change are statistically analyzed to screen for stable optimization schemes. Secondly, the optimization results are independently validated using the first backbone network RiboMamba, which is a pure sequence Disome model that was not involved in the design scoring and has undergone secondary training. This reduces the skepticism introduced by self-validation within the same model and further enhances the credibility of the conclusions. Matching is determined by calculating the correlation coefficient (e.g., Pearson correlation coefficient) or mean squared error between the second and first dimer ribosome density distributions. If the correlation coefficient is greater than a preset threshold (e.g., 0.8) or the mean squared error is less than a preset threshold, the two are considered to match, indicating that the current optimization results are robust. Otherwise, it suggests that the optimization results may have overfitting risks and need to be re-optimized.

[0044] This invention proposes a computational process that includes two-level collision prediction, result calibration, synonymous codon optimization, global detection, and orthogonal verification. The core consists of two models: a first backbone network, RiboMamba, takes a full-length codon sequence as input to predict the monosome density distribution; a second backbone network, RiboMamba, takes a full-length codon sequence and its corresponding monosome density combination as input to predict the disome density distribution; and an independently trained first backbone network, RiboMamba, uses a full-length codon sequence as input to predict the disome density distribution for orthogonal verification, evaluating the generalization ability and reliability of the entire process.

[0045] The first and second backbone networks, RiboMamba and RiboMamba respectively, are two functional models used in this invention. Their networks are built from publicly available deep learning fundamental modules, including embedding, positional encoding, convolutional layers, and Mamba state space blocks. This invention names the custom backbone network formed based on the above module combination RiboMamba. Both models use this RiboMamba backbone network to enable rapid computation and iterative processing of the full-length CDS.

[0046] Based on this, the present invention divides the prediction task into two steps. First, the first backbone network RiboMamba predicts the background density of the Monosome, and then the second backbone network RiboMamba predicts the collision density of the Disome. The calibrated output of the first backbone network RiboMamba is used at the input of the second backbone network RiboMamba, allowing the collision prediction to reference background information. Instead of outputting a single score for a specific center position, the present invention outputs continuous values ​​for each codon position, obtaining the density distribution of the entire gene. This achieves peak reduction within the window and simultaneously checks for newly squeezed peaks outside the window. Additionally, a Disome model RiboMamba is trained that does not participate in the optimization scoring process; it is only used to independently verify the optimization results, thereby improving the reliability of the results. All local peaks in the Monosome density distribution are identified, sorted by peak height, and the locations of the Top-N peaks are selected as optimization centers. Each center is extended outwards by a preset length (e.g., 50 codons) to form an optimization window; or, a target region can be directly preset as the optimization window according to experimental requirements. Using the initial score of the original gene sequence as a benchmark, a relative threshold (such as 50%-80% of the initial score) can be set; or an absolute threshold can be set based on historical optimization data; or an adaptive threshold strategy can be adopted in multiple iterations, and convergence can be determined when the score no longer decreases significantly in multiple consecutive iterations, and the current score can be used as the threshold standard.

[0047] Using the yeast disome / monosome dataset as an example, approximately 1632 genes were modeled; synonym optimization was performed on genes such as YDR172W and YLL024C. Taking the yeast gene YDR172W as an example, as the algorithm iterated according to the scoring objective function, an optimal candidate sequence containing 29 synonymous mutations was generated. The predicted peak value of the disome ribosome density of this sequence within the target optimization window significantly decreased from the initial 1.082 to 0.501, with a local congestion peak reduction of 53.7%. Meanwhile, global health checks using Guardrail confirmed no abnormal peak increases outside the optimization window, and orthogonal validation using the independently trained first backbone network RiboMamba showed a highly consistent optimization trend, fully demonstrating that this sequence optimization method has extremely high global safety and physical robustness.

[0048] This invention offers the following advantages: It constructs an executable sequence design pipeline through full-length density distribution prediction, local optimization, global health checks, orthogonal validation, and sequence derivation; the modeling mechanism is clearly hierarchical, with the first backbone network characterizing the background density of monomeric ribosomes and the second backbone network modeling the collision formation process, making the optimization results and scoring sources interpretable; it supports full-length density distribution output, enabling intuitive monitoring of peak changes in the target window and regions outside the window, meeting the global visualization needs of gene engineering design; it introduces a global linear calibration mechanism to eliminate scale bias between the outputs of the first and second backbone networks, improving the numerical stability of multi-model cascade inference; it sets an out-of-window penalty term and, in conjunction with the global health check mechanism, effectively suppresses the rise of distant new peaks caused by local optimization, enhancing the global robustness of the optimization scheme; the optimization results can be directly output as FASTA format sequences and mutation site lists, facilitating subsequent wet experimental validation and synthetic biology applications.

[0049] This invention employs a first backbone network, RiboMamba, to predict the density distribution of monosomes, and a second backbone network, RiboMamba, to predict the density distribution of disomes based on sequence information and monosome density information. This distinguishes between background translation density changes and collision-related changes, facilitating the setting of optimization objectives and the interpretation of results. This invention focuses on synonymous codon optimization of coding sequences, including not only density distribution prediction but also calibration, candidate generation and scoring, global health checks, and independent verification, forming an executable optimization process. This invention sets an out-of-window penalty term in the optimization objective to suppress the rise of new peaks outside the target window. In the result evaluation stage, Guardrail global health checks are used to statistically analyze changes in the target window, out-of-window changes, and overall distribution changes to screen stable solutions. This invention also introduces an independent disome prediction model (independently trained first backbone network RiboMamba) that did not participate in the optimization scoring for orthogonal verification as supplementary verification evidence. This invention aims to reduce disome collision peaks and related risks, balancing out-of-window impact and result reliability under the constraint of the number of synonym substitutions.

[0050] Based on the same inventive concept, this invention also provides a gene synonymous codon optimization system, comprising: The data acquisition module is used to collect the codon index sequence of the original genes of the target organism.

[0051] A feature extraction module is used to input the codon index sequence into a first backbone network, RiboMamba, and output the single-unit ribosome density distribution of the codon index sequence. The first backbone network, RiboMamba, consists of an input layer, an embedding module, a positional encoding, a multi-scale convolutional block, a Mamba state space block, and an output layer connected in sequence. Based on the peak position or a preset target region in the single-unit ribosome density distribution, an optimal window is determined, and a codon site is randomly selected within the optimal window for synonymous codon replacement, resulting in an updated codon index sequence. The updated codon index sequence is then input into a second backbone network, RiboMamba, and outputs a first two-unit ribosome density distribution. The second backbone network, RiboMamba, includes a fusion module added between the multi-scale convolutional block and the Mamba state space block of the first backbone network, RiboMamba. The first two-unit ribosome density distribution is scored, and when the score meets a preset condition, the current optimal codon index sequence is output. The second two-unit ribosome density distribution of the current optimal codon index sequence is extracted again using the first backbone network, RiboMamba.

[0052] An optimization module is used to confirm the current optimal codon index sequence as the final optimal codon index sequence if the density distribution of the second dimer ribosome matches the density distribution of the first dimer ribosome.

[0053] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile storage, and may also include other hardware required for various operations. The processor reads the corresponding computer program from the non-volatile storage into memory and then runs it to implement the gene synonym codon optimization method provided above.

[0054] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the gene synonym codon optimization method provided above.

[0055] Specific limitations regarding the computational system for the gene synonymous codon optimization method can be found in the limitations of the gene synonymous codon optimization method described above, and will not be repeated here. Each module in the aforementioned gene synonymous codon optimization system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0056] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for optimizing gene synonymous codons, characterized in that, Includes the following steps: Collect the codon index sequence of the original gene of the target organism; The codon index sequence is input into the first backbone network RiboMamba, and the monosomal ribosome density distribution of the codon index sequence is output. The first backbone network RiboMamba consists of an input layer, an embedding module, a positional encoding, a multi-scale convolutional block, a Mamba state space block, and an output layer connected in sequence. Based on the peak position or preset target region in the single ribosome density distribution, an optimal window is determined. A codon site is randomly selected within the optimal window for synonymous codon replacement, resulting in an updated codon index sequence. The updated codon index sequence is input into the second backbone network RiboMamba, which outputs a first dual-ribosome density distribution. The second backbone network RiboMamba includes a fusion module added between the multi-scale convolutional blocks and Mamba state space blocks of the first backbone network RiboMamba. The first dual-ribosome density distribution is scored, and when the score meets preset conditions, the current optimal codon index sequence is output. The second dual-ribosome density distribution of the current optimal codon index sequence is extracted again using the first backbone network RiboMamba. If the density distribution of the second dimer ribosome matches the density distribution of the first dimer ribosome, then the current optimal codon index sequence is confirmed as the final optimal codon index sequence.

2. The gene synonymous codon optimization method according to claim 1, characterized in that, The codon index sequence is input into the first backbone network RiboMamba, and the monosomal ribosome density distribution of the codon index sequence is output. Specifically, this includes: mapping each codon index in the codon index sequence to a multi-dimensional encoding vector by a multi-scale convolutional block, and adding positional encoding information to the encoding vector of each dimension to obtain an initial encoding sequence representation; capturing and mapping the long-range dependency encoding information of the initial encoding sequence representation using Mamba state space blocks to obtain the monosomal ribosome density distribution of the codon index sequence.

3. The gene synonymous codon optimization method according to claim 1, characterized in that, The fusion module includes: a splicing layer, a fully connected layer, and a GELU activation function; it uses the second backbone network RiboMamba to extract the initial encoding sequence representation of the updated codon index sequence; it uses the fusion module of the second backbone network RiboMamba to splice the single ribosome density distribution and the initial encoding sequence representation to obtain the fused encoding features; based on the fused encoding features, it outputs the first dual ribosome density distribution.

4. The gene synonymous codon optimization method according to claim 1, characterized in that, The first backbone network, RiboMamba, is trained using two types of data samples: one is trained independently using monomeric ribosome density data, which predicts and outputs the monomeric ribosome density distribution; the other is trained independently using disosome ribosome density data, which predicts and outputs the second disosome ribosome density distribution. The monomeric and disosome ribosome density data are extracted from the codon index sequence of the original gene of the target organism.

5. The gene synonymous codon optimization method according to claim 1, characterized in that, The scoring based on the density distribution of the first dimer ribosome specifically includes: using the sum of the density values ​​of the first dimer ribosomes at all codon sites within the optimal window as the score.

6. The gene synonymous codon optimization method according to claim 1, characterized in that, The optimal window is determined based on the peak position in the monomer ribosome density distribution or a preset target region, and further includes setting an out-of-window penalty term to suppress the generation of new peaks outside the optimal window.

7. The gene synonymous codon optimization method according to claim 3, characterized in that, Before splicing the monomer ribosome density distribution and the initial coding sequence representation, the method further includes: performing global linear calibration on the monomer ribosome density distribution output by the first backbone network RiboMamba, and splicing the calibrated monomer ribosome density distribution and the initial coding sequence representation.

8. A gene synonymous codon optimization system, characterized in that, include: The data acquisition module is used to collect the codon index sequence of the original genes of the target organism; A feature extraction module is used to input the codon index sequence into a first backbone network, RiboMamba, and output the single-unit ribosome density distribution of the codon index sequence. The first backbone network, RiboMamba, consists of an input layer, an embedding module, a positional encoding, a multi-scale convolutional block, a Mamba state space block, and an output layer connected in sequence. Based on the peak position or a preset target region in the single-unit ribosome density distribution, an optimal window is determined, and a codon site is randomly selected within the optimal window for synonymous codon replacement to obtain an updated codon index sequence. The updated codon index sequence is then input into a second backbone network, RiboMamba, and outputs a first two-unit ribosome density distribution. The second backbone network, RiboMamba, includes a fusion module added between the multi-scale convolutional block and the Mamba state space block of the first backbone network, RiboMamba. The first two-unit ribosome density distribution is scored, and when the score meets a preset condition, the current optimal codon index sequence is output. The second two-unit ribosome density distribution of the current optimal codon index sequence is extracted again using the first backbone network, RiboMamba. An optimization module is used to confirm the current optimal codon index sequence as the final optimal codon index sequence if the density distribution of the second dimer ribosome matches the density distribution of the first dimer ribosome.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 7.

Images