VeloHARMO: a deep learning framework for mRNA coding sequence generation and design via translation velocity harmonization Authors: Bian Bian, Jichen Zhang, Yiming Zhang, Yutaka Saito Published: 2026 ABSTRACT AbstractHeterologous gene expression often results in suboptimal protein yield or activity, potentially influenced by differences in translation dynamics between native and host organisms. Conventional mRNA design approaches based on codon usage frequency can improve expression efficiency. However, codon usage frequency alone does not necessarily reflect codon-level translation velocity. Here, we present VeloHARMO, a computational framework for mRNA coding sequence design based on translation velocity harmonization. VeloHARMO combines organism-specific deep learning models trained on Ribo-seq data with a genetic algorithm to generate synonymous coding sequences that preserve native translation velocity patterns across organisms. Computational evaluations show that VeloHARMO achieves higher prediction accuracy and more effective translation velocity harmonization than existing approaches. Furthermore, analyses of previously reported synonymous mRNA variants suggest that translation velocity harmonization may help explain differences observed among alternative coding sequences. Collectively, VeloHARMO introduces translation velocity harmonization as a data-driven principle for computational mRNA design and provides a framework for incorporating organism-specific translation dynamics into coding sequence engineering. Introduction Heterologous gene expression, in which a gene from one organism is expressed in another, is a foundational strategy in biotechnology and biopharmaceuticals [1, 2]. Common host organisms such as Escherichia coli and yeast are widely used due to their fast growth, ease of genetic manipulation, and low production cost [3–5]. However, transferring a gene from its native context to a new host often leads to reduced protein yield or activity [6, 7]. While multiple factors contribute to this, differences in translation dynamics between the native and host organisms are likely to play a role [8], particularly in cases where co-translational protein folding is essential for function [9–11]. Computational approaches have been increasingly applied to RNA sequence analysis and prediction across diverse RNA classes [12]. In parallel, computational mRNA coding sequence design has become increasingly important in vaccine and therapeutic development, including reverse vaccinology-based mRNA vaccine design against emerging viral pathogens [13]. To improve translation efficiency, codon optimization has been widely applied [14]. This method replaces synonymous codons with those more frequently used in the host organism, aiming to enhance expression. However, codon optimization can disrupt the timing of translation, leading to misfolding and aggregation of the synthesized protein [7, 15]. Codon harmonization has emerged as an alternative strategy [16–18], substituting codons based on their usage frequencies to better match the native codon usage frequency pattern along the sequence. Both approaches, however, rely on codon usage frequency as a proxy for translation velocity, which does not account for additional determinants such as tRNA availability, mRNA secondary structure, or regulatory motifs. More recently, deep learning and heuristic search-based frameworks have been developed for mRNA and codon sequence optimization. For example, CodonTransformer uses a transformer-based architecture for multi-species codon optimization, whereas RNASTOP combines deep learning with heuristic search to optimize mRNA chemical stability and translation-related features [19, 20]. Advances in ribosome profiling (Ribo-seq) have enabled codon-resolution characterization of ribosome occupancy and translation dynamics [21–23]. Leveraging these data, recent computational models have been developed to predict translation velocity using sequence features [24–27]. However, existing models often ignore specific sequence positions or do not fully incorporate protein-related features [28] that may influence translation elongation and folding. Here, we introduce VeloHARMO, a computational mRNA design framework based on translation velocity harmonization. VeloHARMO combines organism-specific deep learning models trained on Ribo-seq data with a genetic algorithm to generate synonymous mRNA coding sequences. The method aims to design sequences whose predicted translation velocity profiles in a heterologous host resemble the predicted native translation velocity pattern of the original gene. We evaluate VeloHARMO through codon-level translation velocity prediction, in silico cross-organism mRNA design, and analyses of published synonymous mRNA expression data. Our results suggest that translation velocity harmonization provides a data-driven strategy for incorporating organism-specific translation dynamics into computational mRNA sequence design. Materials and methods Data collection and Preprocessing We used publicly available ribosome profiling (Ribo-seq) datasets from five organisms: E. coli, Saccharomyces cerevisiae (yeast), Drosophila melanogaster (fruit fly), Pseudomonas aeruginosa, and human [29–33]. For each organism, we selected high-quality datasets with sufficient read coverage to support codon-level analysis. Ribo-seq reads were aligned to reference coding sequences, and footprint densities were calculated for each codon position. Positions with zero footprint density were retained in the dataset to preserve biologically meaningful pausing information. For each codon, we extracted input features including the local codon sequence (e.g. surrounding window of ±10 codons), amino acid sequence, and predicted protein secondary structure. Secondary structure was predicted using DSSP tools from the amino acid sequence. All features were encoded as one-hot vectors or token embeddings as appropriate. Deep learning models for translation velocity prediction We developed two deep learning models, VeloGRU and VeloBERT, to predict codon-level translation velocity from sequence features. VeloGRU is a recurrent neural network based on GRUs, which captures long-range dependencies in codon and amino acid sequences. Input features include codon sequence, amino acid sequence, and secondary structure. The model outputs a continuous value representing normalized ribosome occupancy for each codon. VeloBERT is a transformer-based model built upon a pretrained DNABERT backbone [34] for codon-level translation velocity prediction. The pretrained DNABERT encoder was fine-tuned using organism-specific Ribo-seq data to generate contextual codon representations, which were subsequently fused with protein-derived features, including amino acid sequence and protein secondary structure. The fused representations were then passed to a regression head to predict normalized ribosome occupancy for each codon. Both models were trained using mean squared error loss and evaluated using ten-fold cross-validation. Performance was measured in terms of R2 between predicted and observed translation velocity. Genetic algorithm for translation velocity harmonization To design synonymous mRNA sequences with harmonized translation velocity, we developed a genetic algorithm that iteratively optimizes synonymous codon substitutions. Given a protein-coding sequence from a native organism, the algorithm searches for a synonymous coding sequence in the host organism whose predicted translation velocity profile closely matches the native profile. The optimization objective is to maximize the Pearson correlation between the predicted translation velocity profile of the redesigned sequence in the host organism and the predicted translation velocity profile of the original sequence in the native organism. Each candidate sequence is initialized by random synonymous substitutions while preserving the encoded amino acid sequence. At each generation, candidate sequences are evaluated using this fitness score. High-fitness sequences are selected and used to generate the next population through crossover and synonymous codon mutation. Unless otherwise specified, the genetic algorithm was performed with a population size of 100, mutation rate of 0.01, and 1000 generations. We applied this algorithm using VeloGRU models trained separately for the source and target organisms. To assess the stability of the genetic algorithm, we performed additional analyses using P68767 (pepA) as a representative gene. We repeated the optimization using three different random seeds and recorded the best fitness score at each generation to evaluate convergence behavior. We also conducted a parameter sensitivity analysis by varying one parameter at a time, including generation number, mutation rate, and population size. Runtime, initial fitness, final fitness, and CAI values of the designed sequences were further calculated to evaluate computational efficiency and codon usage properties. Benchmarking and evaluation To evaluate prediction performance, we compared VeloGRU and VeloBERT with a previously published model using Ribo-seq data from five organisms. We also tested the generalizability of models across organisms by cross-predicting between E. coli and yeast. For mRNA design evaluation, we applied VeloHARMO to 3808 E. coli genes for expression in yeast. For each gene, we compared the predicted translation velocity correlation before and after redesign and calculated the fitness change relative to the wild-type sequence. We benchmarked VeloHARMO against CHARMING, CodonTransformer, conventional codon optimization, codon usage frequency-matched random synonymous substitution, and tAI-based harmonization. These comparisons were used to evaluate whether velocity-guided optimization improves translation velocity harmonization beyond codon usage composition, codon optimization, tRNA adaptation, random synonymous substitution, and deep learning-based codon design. Additional validation was performed on previously published synonymous mRNA sequences encoding firefly luciferase, PTP4A3, and iLRP. Fitness scores were computed as the correlation between the predicted translation velocity profile of the redesigned sequence and the original native sequence. The fitness score therefore reflects the similarity between the predicted translation velocity profiles rather than a direct measure of protein expression or translation efficiency, with higher values indicating greater profile similarity. Results Overview of VeloHARMO VeloHARMO is a computational framework for designing synonymous mRNA that preserve translation velocity pattern along a sequence between native and heterologous host organisms (Fig. 1). The method is based on the concept of translation velocity harmonization, which aims to match the codon-level translation velocity of a gene in its native organism when expressed in a different host. Framework of VeloHARMO. (A) Deep learning modeling for translation velocity prediction. (B) The algorithm used for mRNA coding sequence harmonization. VeloHARMO consists of two core components: (i) deep learning models trained on Ribo-seq data to predict translation velocity pattern from sequence and protein features (Fig. S1 and Fig. S2) and (ii) a genetic algorithm that optimizes synonymous codons to reproduce the native translation velocity pattern in the host organism (Fig. S3). This framework allows for organism-specific modeling while accounting for complex determinants of translation velocity beyond codon usage frequency, including amino acid sequence and protein secondary structure features. Deep learning models predict codon-level translation velocity from sequence and protein structural features To model translation velocity pattern from sequence data, we developed two deep learning models: VeloGRU (Fig. S1) and VeloBERT (Fig. S2). Both models were trained using Ribo-seq data to predict codon-level translation velocity, integrating codon sequence, amino acid sequence, and protein secondary structure features. VeloGRU, based on a gated recurrent unit (GRU), was designed to capture long-range sequential dependencies in translation. VeloBERT is built upon a fine-tuned pretrained DNABERT backbone [34], where contextual codon representations are integrated with protein-derived features, including amino acid sequence and protein secondary structure, for codon-level translation velocity prediction (Fig. S2). We trained organism-specific models using codon-wise Ribo-seq footprint density as a proxy for translation velocity. Model performance was evaluated on E. coli, yeast, fruit fly, Pseudomonas aeruginosa, and human datasets using 10-fold cross-validation. In both E. coli and yeast, VeloGRU achieved higher accuracy than a previously published model, RiboMIMO [25] (Table 1). We also evaluated the inclusion of protein secondary structure features motivated by the relationship between translation elongation and cotranslational protein folding. Although secondary structure information did not consistently improve overall prediction accuracy across datasets, it was retained as a biologically relevant feature for investigating potential associations between translation dynamics and protein structural properties (Table 1 and Table 2). VeloBERT exhibited superior performance over VeloGRU in smaller datasets such as P. aeruginosa (Table 2), potentially due to the benefit of transfer learning from the human-genome-pretrained DNABERT. Table 1Performance evaluation of RiboMIMO and VeloGRU, measured in terms of the R2 between the true translation velocity and predicted translation velocity.Methods E. coli E. coli (ss)YeastYeast (ss)RiboMIMO0.368 ± 0.040.375 ± 0.040.400 ± 0.020.397 ± 0.01VeloGRU0.429 ± 0.02(***)0.423 ± 0.02(***)0.459 ± 0.01(***)0.454 ± 0.01(***)10-fold cross-validation through randomly splitting genes into training and test set on E. coli and yeast dataset to assess the performance of each model. The mean ± SD of the R2 are shown. For each fold, R2 was calculated for each gene along a sequence, then the average value for all genes. Ss stand for protein secondary structure elements. Student’s t-test was used, *P < 0.05, **P < 0.01, ***P < 0.001. Performance evaluation of RiboMIMO and VeloGRU, measured in terms of the R2 between the true translation velocity and predicted translation velocity. 10-fold cross-validation through randomly splitting genes into training and test set on E. coli and yeast dataset to assess the performance of each model. The mean ± SD of the R2 are shown. For each fold, R2 was calculated for each gene along a sequence, then the average value for all genes. Ss stand for protein secondary structure elements. Student’s t-test was used, *P < 0.05, **P < 0.01, ***P < 0.001. Table 2Performance evaluation of VeloGRU and VeloBERT for E. coli, yeast, fruit fly, P. aeruginosa and human dataset, measured in terms of the R2 between the true translation velocity and predicted translation velocity.AMethods E. coli YeastFruit fly P. aeruginosa HumanVeloGRU0.429 ± 0.02 *** 0.459 ± 0.01 *** 0.408 ± 0.020.094 ± 0.010.281 ± 0.01VeloBERT0.427 ± 0.010.435 ± 0.010.369 ± 0.02 ** 0.13 ± 0.03 ** 0.313 ± 0.02 B Methods E. coli (ss) Yeast (ss) Fruit fly (ss) P. aeruginosa (ss) Human (ss) VeloGRU0.423 ± 0.02 ** 0.454 ± 0.01 *** 0.406 ± 0.010.085 ± 0.010.284 ± 0.02VeloBERT0.433 ± 0.010.436 ± 0.010.363 ± 0.02 *** 0.129 ± 0.02 ** 0.315 ± 0.03(A) Performance evaluation without incorporating protein secondary structure elements in each dataset. (B) Performance evaluation incorporating protein secondary structure elements in each dataset. 10-fold cross validation through randomly splitting genes into training and test set to assess the performance of each method in each dataset. Ss stand for protein secondary structure elements. Student’s t-test was used, *P < 0.05, **P < 0.01, ***P < 0.001. Performance evaluation of VeloGRU and VeloBERT for E. coli, yeast, fruit fly, P. aeruginosa and human dataset, measured in terms of the R2 between the true translation velocity and predicted translation velocity. (A) Performance evaluation without incorporating protein secondary structure elements in each dataset. (B) Performance evaluation incorporating protein secondary structure elements in each dataset. 10-fold cross validation through randomly splitting genes into training and test set to assess the performance of each method in each dataset. Ss stand for protein secondary structure elements. Student’s t-test was used, *P < 0.05, **P < 0.01, ***P < 0.001. Translation velocity profiles differ significantly between organisms To assess whether translation velocity patterns are conserved across organisms, we performed cross-species prediction experiments. A VeloGRU model trained on E. coli data was applied to yeast sequences and vice versa (Fig. S4). Both experiments resulted in low predictive performance, indicating a lack of generalizability between species. This suggests that translation velocity is strongly organism-specific and supports the need for species-specific modeling in mRNA design. VeloHARMO designs mRNA sequences by harmonizing translation velocity patterns We implemented a genetic algorithm that uses organism-specific VeloGRU models to design synonymous mRNA sequences that preserve the native translation velocity pattern when expressed in a heterologous host (Fig. S3). The algorithm iteratively mutates and selects synonymous codon variants to maximize the Pearson correlation between the predicted velocity profile in the host and the native organism. To validate this approach, we first applied VeloHARMO to a toy problem within yeast (both the host and native organisms set as yeast). This correctly resulted in an mRNA sequence reproducing the native translation velocity pattern starting from an initial random sequence (Fig. S5, Pearson’s correlation coefficient = 0.995). We then applied VeloHARMO to design mRNA sequences of 3808 E. coli genes for expression in yeast (Fig. S3). In nearly all cases, the harmonized sequences showed improved correlation with the native velocity pattern, indicating successful translation velocity harmonization (Fig. 2). mRNA sequence design by VeloHARMO and comparison to CHARMING method. (A) mRNA sequence design by VeloHARMO for a highly expressed gene. Population size is 100, mutation rate is 0.01, generation number is 1000 for VeloHARMO, and final Pearson correlation is 0.43. (B) Comparative analysis of fitness change between VeloHARMO and codon harmonization by CHARMING. To further assess the robustness of the genetic algorithm, we performed additional analyses using P68767 (pepA) as a representative gene. Three independent runs with different random seeds produced similar convergence trajectories (Fig. S6). The runtime and best fitness score were also comparable across runs, and the optimized sequences showed higher CAI values than the wild-type sequence (Table S1), suggesting that VeloHARMO-designed sequences remain compatible with host codon usage preferences. We further performed parameter sensitivity analysis by varying generation number, mutation rate, and population size. The default setting showed comparable performance to longer runs and larger population settings, while a higher mutation rate resulted in lower fitness (Table S2). These results indicate that the genetic algorithm was reasonably stable under different random initializations. VeloHARMO improves upon existing mRNA design methods To benchmark VeloHARMO against existing approaches, we firstly compared its harmonization performance with a previously published codon harmonization method, CHARMING [18]. For each E. coli gene designed for yeast expression, we computed the change in velocity correlation (‘fitness’) before and after the design. VeloHARMO consistently yielded greater fitness improvement across genes, whereas codon harmonization (CHARMING) methods showed only limited enhancement (Fig. 2B). This result suggests that matching codon usage patterns alone does not reliably ensure translation velocity harmonization. We further expanded the benchmark by comparing VeloHARMO with additional baseline methods, including CodonTransformer, conventional codon optimization, codon usage frequency-matched random synonymous substitution, and tAI-based harmonization. Across 3808 E. coli genes, these additional baselines also showed only limited fitness improvement compared with VeloHARMO (Fig. S7). These results indicate that translation velocity harmonization cannot be achieved simply by matching codon usage composition, optimizing host-preferred codons, preserving tRNA adaptation patterns, or applying a general deep learning-based codon design method. Instead, the improved fitness achieved by VeloHARMO indicates that directly using organism-specific translation velocity information can better guide synonymous mRNA sequence design. VeloHARMO fitness is consistent with reported expression differences among synonymous mRNA sequences We further evaluated the utility of VeloHARMO by analyzing previously published synonymous mRNA sequences designed using codon or tRNA usage–based methods. These sequences were originally optimized for heterologous expression of firefly luciferase [16], HPDF, PTP4A3, and iLRP [4, 35] in E. coli. Using our trained translation velocity prediction models, we computed fitness scores, defined as the velocity profile correlation between the native and host contexts, for the wild-type sequences, previously published optimized sequences, and VeloHARMO-designed sequences. Although the number of available sequences with quantitative experimental measurements was limited and did not allow robust statistical correlation analysis, the reported experimental trends were generally consistent with the fitness scores. For luciferase, PTP4A3, and iLRP, sequences associated with improved activity or expression showed higher fitness than the corresponding wild-type sequences, and VeloHARMO-designed sequences achieved the highest predicted fitness scores (Table 3). These results provide qualitative support that translation velocity harmonization may help interpret expression differences among synonymous coding sequences, while larger quantitative datasets will be needed for rigorous experimental validation. Table 3Fitness evaluation of publicly available mRNA sequence.AGene (protein)Experimental ObservationWT FitnessPublished Optimized FitnessVeloHARMO FitnessLuciferaseActivity increased−0.0150.0580.453BGene (protein)Experimental observationWT fitnessPublished optimized fitnessVeloHARMO fitnessHPDFNA0.087−0.0120.661PTP4A3Expression increased−0.146−0.1400.374iLRPStronger western blot band−0.085−0.0620.456(A) Fitness score (translation velocity correlation between fruit fly and E. coli) of luciferase wild-type mRNA sequence, the mRNA sequence generated from t-RNA usage frequency-based harmonization and VeloHARMO method. Population size is 100, mutation rate is 0.01, generation number is 1000 for VeloHARMO. (B) Fitness score (translation velocity correlation between human and E. coli) of wild-type mRNA sequences of each gene, mRNA sequences generated from codon optimization and VeloHARMO method. Population size is 100, mutation rate is 0.01, generation number is 1000 for VeloHARMO. NA for HPDF indicates that the corresponding sequence was not experimentally evaluated in the previous study. Fitness evaluation of publicly available mRNA sequence. (A) Fitness score (translation velocity correlation between fruit fly and E. coli) of luciferase wild-type mRNA sequence, the mRNA sequence generated from t-RNA usage frequency-based harmonization and VeloHARMO method. Population size is 100, mutation rate is 0.01, generation number is 1000 for VeloHARMO. (B) Fitness score (translation velocity correlation between human and E. coli) of wild-type mRNA sequences of each gene, mRNA sequences generated from codon optimization and VeloHARMO method. Population size is 100, mutation rate is 0.01, generation number is 1000 for VeloHARMO. NA for HPDF indicates that the corresponding sequence was not experimentally evaluated in the previous study. Discussion Substituting synonymous codons alters the mRNA sequence without changing the encoded amino acid sequence [15]. Over recent decades, it has become increasingly evident that such substitutions can modulate the expression levels of recombinant proteins [36, 37]. Accordingly, codon optimization and harmonization have been widely adopted in heterologous expression systems [38–41]. However, these approaches rely heavily on codon usage frequency as a surrogate for translation velocity and often overlook other determinants of elongation dynamics and protein folding. In this study, we proposed translation velocity harmonization as a computational principle for mRNA coding sequence design and developed VeloHARMO, a deep learning-based framework that implements this concept using organism-specific translation velocity models trained on Ribo-seq data. Although previous codon harmonization studies have attempted to modulate translation elongation rates using codon usage or tRNA usage information [18], VeloHARMO differs by directly modeling codon-level translation velocity and optimizing synonymous sequences according to predicted native and host velocity profiles. It then applies a genetic algorithm to design synonymous mRNA sequences that preserve the native translation velocity pattern in a heterologous host. The method is applicable to any organism pair with available Ribo-seq data, potentially enabling broader application beyond traditional model hosts. We explored two deep learning architectures for translation velocity prediction: the recurrent VeloGRU and the attention-based VeloBERT. While previous models such as RiboMIMO, iXnos, and ROSE employed convolutional or recurrent neural networks [24–26], we introduced DNABERT, a pre-trained transformer model, into codon-level translation velocity prediction. The comparison showed that VeloBERT performed better than VeloGRU in the human and P. aeruginosa datasets, whereas VeloGRU achieved higher accuracy in E. coli, yeast, and fruit fly. This suggests that the benefit of VeloBERT is organism and dataset-dependent. Because DNABERT was pretrained on the human genome, its representations may not fully generalize to bacterial, fungal, or insect coding sequences. Therefore, VeloBERT should be regarded as an exploratory transfer learning-based model rather than a universally superior architecture. Future work using organism-specific or multi-species pretrained nucleotide models may help reduce this domain mismatch and improve the applicability of transformer-based translation velocity prediction across diverse heterologous expression systems. We also evaluated the contribution of protein structural information to translation velocity prediction. Ribosome pausing has been reported to occur near structural boundaries and may contribute to co-translational protein folding [24]. Motivated by this biological relationship, we integrated protein secondary structure features into both VeloGRU and VeloBERT. However, the inclusion of secondary structure information did not consistently improve overall prediction performance across datasets (Table 2). These results suggest that secondary structure information alone is insufficient to substantially improve global prediction accuracy in the current setting. Its contribution to predictive performance may therefore be organism and model dependent. Nevertheless, protein structural information may still provide biologically relevant context for understanding translation dynamics. Further investigation using larger datasets and more detailed structural annotations will be required to clarify its role. Our results demonstrate that codon usage frequency is not a reliable proxy for translation velocity. Although codon harmonization has been proposed to address limitations of codon optimization,particularly for the correct folding of membrane and aggregation-prone proteins [41, 42]. Our findings suggest that codon usage similarity does not guarantee harmonized translation dynamics. For instance, translation velocity profiles differ substantially between E. coli and yeast (Figs S3 and S4), and harmonization based on codon usage alone did not produce similar velocity patterns (Fig. 2B). These results may help explain why codon optimization occasionally fails to produce functional proteins despite high expression. Codon-level translation velocity is influenced by multiple factors beyond the local coding sequence features used in the current model. These factors may include tRNA abundance and charging levels, ribosome traffic and collisions, mRNA secondary structure, RNA modifications, transcript-specific regulatory elements, experimental noise in Ribo-seq measurements, and condition-dependent cellular states. In addition, Ribo-seq-derived ribosome occupancy is an indirect proxy for translation velocity and can be affected by technical biases in library preparation, read alignment, footprint assignment, and normalization. Therefore, the predicted translation velocity profiles in VeloHARMO should be interpreted as Ribo-seq-derived estimates of relative translation dynamics rather than direct measurements of absolute translation elongation rates. These biological and technical sources of variability likely contribute to the unexplained variance. While most codon optimization studies have focused on E. coli and yeast due to their fast growth and well-characterized biology [43, 44], emerging applications in synthetic biology and biotechnology demand tools that can support a wider range of host organisms. VeloHARMO is broadly applicable, provided that Ribo-seq data are available, and may assist in selecting host organisms based on their translational compatibility with a given gene. As more Ribo-seq datasets become available for non-model organisms, the scope of translation velocity harmonization will likely expand. The parameter sensitivity analysis further suggests that the default genetic algorithm settings provide a practical balance between optimization performance and computational cost rather than universally optimal parameters. Larger populations or additional generations may be considered when broader exploration of the synonymous sequence space is desired, at the cost of increased computation, whereas excessively high mutation rates may reduce optimization stability. Therefore, parameter selection may be adjusted according to the specific design task and available computational resources. Finally, our analysis of previously reported synonymous coding sequence variants suggests that VeloHARMO fitness is qualitatively consistent with experimentally observed expression differences in several published examples. Although these retrospective comparisons do not constitute direct experimental validation and are limited by the small number of available sequences, they indicate that translation velocity harmonization may provide a useful computational perspective for interpreting expression differences among synonymous mRNA sequences. In summary, VeloHARMO provides a data-driven computational framework for mRNA coding sequence design by incorporating organism-specific translation dynamics into synonymous sequence optimization. The current study establishes translation velocity harmonization as an in silico design principle and highlights its potential relevance for heterologous gene expression. Future experimental validation using quantitatively measured expression, activity, solubility, and folding outcomes will be required to evaluate the practical utility of VeloHARMO-designed sequences. Key points We propose VeloHARMO, a deep learning and genetic algorithm-based framework for mRNA sequence design via translation velocity harmonization. We propose VeloHARMO, a deep learning and genetic algorithm-based framework for mRNA sequence design via translation velocity harmonization. VeloHARMO learns organism-specific codon-level translation velocity from Ribo-seq data and generates synonymous coding sequences that mimic native translation dynamics. VeloHARMO learns organism-specific codon-level translation velocity from Ribo-seq data and generates synonymous coding sequences that mimic native translation dynamics. VeloHARMO achieves improved translation velocity prediction and harmonization performance compared with existing mRNA design benchmarks. VeloHARMO achieves improved translation velocity prediction and harmonization performance compared with existing mRNA design benchmarks. VeloHARMO establishes a computational framework for evaluating and designing synonymous mRNA sequences through organism-specific translation velocity modeling. VeloHARMO establishes a computational framework for evaluating and designing synonymous mRNA sequences through organism-specific translation velocity modeling. Supplementary Material