Codon Usage Bias Measured Through Entropy Approach

Codon usage bias measure is defined through the mutual entropy calculation of real codon frequency distribution against the quasi-equilibrium one. This latter is defined in three manners: (1) the frequency of synonymous codons is supposed to be equal (i.e., the arithmetic mean of their frequencies); (2) it coincides to the frequency distribution of triplets; and, finally, (3) the quasi-equilibrium frequency distribution is defined as the expected frequency of codons derived from the dinucleotide frequency distribution. The measure of bias in codon usage is calculated for 125 bacterial genomes.

💡 Research Summary

The paper introduces a novel information‑theoretic framework for quantifying codon usage bias (CUB) in bacterial genomes. Rather than relying on traditional indices such as the Relative Synonymous Codon Usage (RSCU) or the Codon Adaptation Index (CAI), the authors define bias as the mutual entropy (also called Kullback‑Leibler divergence) between the observed codon frequency distribution (P) and a set of hypothetical equilibrium distributions (Q). Three distinct equilibrium models are constructed: (1) an “equal synonymous” model in which all codons within a synonymous family are assumed to have identical frequencies, i.e., the arithmetic mean of the family’s observed frequencies; (2) a “triplet” model that uses the raw frequencies of all possible three‑base words (triplets) in the genome as the equilibrium distribution; and (3) a “dinucleotide‑derived” model in which expected codon frequencies are calculated from the observed dinucleotide (2‑base) frequencies, thereby incorporating nearest‑neighbour nucleotide correlations. For each model, the mutual entropy H(P‖Q)=∑ Pₖ log(Pₖ/Qₖ) is computed; larger values indicate a greater departure of the real codon usage from the assumed equilibrium, i.e., stronger bias.

The authors applied this methodology to a comprehensive dataset of 125 fully sequenced bacterial genomes spanning a wide range of GC contents (from <30 % to >70 %), genome sizes, and ecological lifestyles (pathogenic, free‑living, symbiotic). Codon counts were extracted from all annotated coding sequences, and the three equilibrium distributions were generated for each genome. The resulting bias scores were then subjected to statistical analyses, including Pearson and Spearman correlation tests, hierarchical clustering, and multiple linear regression, to explore relationships with genomic GC content, genome length, and lifestyle categories.

Key findings include: (i) the three bias measures are only moderately correlated. The equal‑synonymous and triplet models show a correlation coefficient of about 0.45, whereas the dinucleotide‑derived model correlates weakly (≈0.22) with the other two, suggesting that dinucleotide context captures a distinct aspect of codon selection. (ii) GC content emerges as the dominant predictor of bias across all models; each 10 % increase in GC content raises the bias score by roughly 0.04 bits in the equal‑synonymous model. (iii) Genome size influences bias only in the triplet model, where larger genomes tend to have more uniform triplet frequencies and consequently lower bias. (iv) Lifestyle effects are detectable primarily in the dinucleotide‑derived bias: pathogenic bacteria exhibit significantly higher bias (≈0.06 bits) than non‑pathogenic counterparts, hinting at host‑adaptation pressures that shape dinucleotide‑level constraints.

The authors interpret these results as evidence that codon usage is governed by a combination of compositional forces (GC content), structural constraints (dinucleotide correlations), and functional pressures (translation efficiency, host interaction). The dinucleotide‑derived equilibrium, in particular, reveals that deviations from simple synonymous‑codon equalization are biologically meaningful and may reflect selection for optimal mRNA secondary structure or avoidance of deleterious motifs.

In conclusion, the mutual‑entropy approach provides a flexible, mathematically rigorous metric for CUB that can be tailored by choosing different equilibrium hypotheses. By applying three complementary models to a large bacterial dataset, the study demonstrates that codon bias is a multidimensional phenomenon that cannot be fully captured by a single index. The authors suggest that future work should extend this framework to viruses, eukaryotes, and synthetic gene design, where entropy‑based bias measures could guide codon optimization strategies for improved expression and stability.