Codon optimization modifies a gene’s coding sequence to enhance
protein production without changing the encoded amino acid sequence. The
codon_optimize function in the cubar package
provides three strategies for optimizing coding sequences based on the
codon usage of the target organism. The first two strategies replace
rare codons with more frequently used ones, while the third strategy
employs the third-party deep learning model
CodonTransformer (Fallahpour et al., 2025) to optimize
codon usage.
Additionally, cubar can integrate the state-of-the-art
deep learning model SpliceAI (Jaganathan et al., 2019) to
prevent the unintended introduction of cryptic splice sites in optimized
sequences. As both SpliceAI and
CodonTransformer are Python-based, users must manually
install these packages. Here we demonstrates how to install them using
conda (or mamba) in a new environment for use
in cubar.
# create a new environment named "cubar_env" with both python and r installed
conda create -n cubar_env python=3.12 r-base blas=*=netlib r-reticulate
# activate the environment we just created
conda activate cubar_env
# install CodonTransformer and SpliceAI
pip install CodonTransformer tensorflow spliceaiThe default “naive” method simply replaces each codon to the most preferred one in the same family or subfamily.
library(cubar)
seq <- 'ATGCTACGA'
cf_all <- count_codons(yeast_cds)
#> Loading required package: Biostrings
#> Warning: package 'Biostrings' was built under R version 4.3.3
#> Loading required package: BiocGenerics
#>
#> Attaching package: 'BiocGenerics'
#> The following objects are masked from 'package:stats':
#>
#> IQR, mad, sd, var, xtabs
#> The following objects are masked from 'package:base':
#>
#> Filter, Find, Map, Position, Reduce, anyDuplicated, aperm, append,
#> as.data.frame, basename, cbind, colnames, dirname, do.call,
#> duplicated, eval, evalq, get, grep, grepl, intersect, is.unsorted,
#> lapply, mapply, match, mget, order, paste, pmax, pmax.int, pmin,
#> pmin.int, rank, rbind, rownames, sapply, setdiff, sort, table,
#> tapply, union, unique, unsplit, which.max, which.min
#> Loading required package: S4Vectors
#> Loading required package: stats4
#>
#> Attaching package: 'S4Vectors'
#> The following object is masked from 'package:utils':
#>
#> findMatches
#> The following objects are masked from 'package:base':
#>
#> I, expand.grid, unname
#> Loading required package: IRanges
#> Loading required package: XVector
#> Loading required package: GenomeInfoDb
#> Warning: package 'GenomeInfoDb' was built under R version 4.3.3
#>
#> Attaching package: 'Biostrings'
#> The following object is masked from 'package:base':
#>
#> strsplit
optimal_codons <- est_optimal_codons(cf_all)
seq_opt <- codon_optimize(seq, optimal_codons)
print(seq_opt)
#> 9-letter DNAString object
#> seq: ATGCTACGTThe “IDT” option implements the method used by the codon optimization tool of Integrated DNA Technologies. Briefly, this method randomly selects synonymous codons from the same family or subfamily based on their relative frequency, but excluding rare codons used below 10% in the target organism.
The “CodonTransformer” method optimizes codon usage with the
third-party software CodonTransformer directly using a
wrapper in R. CodonTransformer is a deep learning model
that can generate coding sequences that show similar codon usage and
distribution to host genes with reduced negative cis elements in a wide
range of organisms across the tree of life. Please refer to the original
study for more details.
seq_opt <- codon_optimize(seq, method = "CodonTransformer", organism = "Saccharomyces cerevisiae")
print(seq_opt)
#> 9-letter DNAString object
#> seq: ATGTTAAGATGAcubar can generate several optimized sequences at the
same time using the argument num_sequences with the method
“IDT” and “CodonTransformer”. When num_sequences is greater
than 1, identical duplicate sequences will be retained as a single copy,
potentially resulting in a final sequence count less than the specified
value.
seqs_opt <- codon_optimize(seq, cf = cf_all, method = "IDT", num_sequences = 10)
print(seqs_opt)
#> DNAStringSet object of length 6:
#> width seq
#> [1] 9 ATGCTCCGT
#> [2] 9 ATGCTGCGT
#> [3] 9 ATGCTTCGT
#> [4] 9 ATGCTACGT
#> [5] 9 ATGCTCCGA
#> [6] 9 ATGCTTCGA
seqs_opt <- codon_optimize(seq, method = "CodonTransformer", organism = "Saccharomyces cerevisiae",
num_sequences = 10, deterministic =FALSE, temperature = 0.4)
print(seqs_opt)
#> DNAStringSet object of length 4:
#> width seq
#> [1] 12 ATGTTGAGATAA
#> [2] 12 ATGTTAAGATAA
#> [3] 12 ATGTTGAGATGA
#> [4] 12 ATGTTGAGATAGIn addition, cubar integrated the deep learning tool
SpliceAI to identify potential splice sites with the
argument spliceai. When the probabilities of non-splice
site for each base are greater than 0.5, it is considered that there are
no potential splice junction sites, and the
Possible_splice_junction in the output is marked as FALSE,
otherwise it is marked as TRUE.
seqs_opt <- codon_optimize(seq, cf = cf_all, method = "IDT", num_sequences = 10, spliceai = TRUE)
print(seqs_opt)
#> Candidate_optimized_sequence Possible_splice_junction
#> <char> <lgcl>
#> 1: ATGCTACGC FALSE
#> 2: ATGCTGCGA FALSE
#> 3: ATGCTGCGT FALSE
#> 4: ATGCTTCGC FALSE
#> 5: ATGCTACGT FALSE
#> 6: ATGCTTCGG FALSE
#> 7: ATGCTCCGT FALSE
#> 8: ATGCTTCGT FALSE
#> 9: ATGCTCCGA FALSE
#> 10: ATGCTTCGA FALSE
seq_opt <- codon_optimize(seq, method = "CodonTransformer", organism = "Saccharomyces cerevisiae", spliceai = TRUE)
print(seq_opt)
#> Candidate_optimized_sequence Possible_splice_junction
#> <char> <lgcl>
#> 1: ATGTTAAGATGA FALSECodon usage within a coding sequence influences multiple aspects of
mRNA biology, including RNA secondary structure, translation elongation,
co-translational folding, and mRNA stability (Liu et al., 2021).
Achieving an optimal coding sequence requires carefully balancing all of
these factors, making codon optimization a more complex task than it may
initially appear. The naive approach implemented in
cubar—which simply replaces each non-optimal codon with the most
preferred one—fails to account for local sequence context and may lead
to unintended consequences. For instance, rare codons are often
strategically positioned to facilitate proper folding of protein
domains; indiscriminately replacing them with preferred codons can
disrupt this process and promote aggregation (Moss et al., 2024).
Given these complexities, we recommend using more sophisticated and context-aware methods for codon optimization. Established strategies such as that implemented in the IDT’s codon optimization tool have demonstrated long-term effectiveness and are widely adopted within the research community. Meanwhile, newer approaches based on deep generative models—such as CodonTransformer (Fallahpour et al., 2025) and CodonBert (Li et al., 2024)—leverage large-scale natural sequence data and advanced architectures like recurrent neural networks and attention mechanisms to capture context-dependent codon usage patterns (Novakovsky et al., 2023). These deep learning–based models offer a powerful, flexible framework for codon optimization and are likely to play a central role in future sequence design workflows.