Abstract
Synonymous codons translate into chemically identical amino acids. Once considered inconsequential to the formation of the protein product, there is evidence to suggest that codon usage affects cotranslational protein folding and the final structure of the expressed protein. Here we develop a method for computing and comparing codonspecific Ramachandran plots and demonstrate that the backbone dihedral angle distributions of some synonymous codons are distinguishable with statistical significance for some secondary structures. This shows that there exists a dependence between codon identity and backbone torsion of the translated amino acid. Although these findings cannot pinpoint the causal direction of this dependence, we discuss the vast biological implications should coding be shown to directly shape protein conformation and demonstrate the usefulness of this method as a tool for probing associations between codon usage and protein structure. Finally, we urge for the inclusion of exact genetic information into structural databases.
Similar content being viewed by others
Introduction
One of the most critical cellular processes is the decoding of genetic information into functional proteins. Transfer RNA (tRNA) molecules recognize codons of the messenger RNA (mRNA) sequence as it passes through the ribosome and deliver specific amino acids sequentially for addition to the growing peptide chain. 61 codons map to 20 amino acids, meaning that most amino acids are encoded by more than one, synonymous, codon. Once considered a silent redundancy of the genetic code, synonymous coding is now known to be functionally important, subject to evolutionary selective pressure and clearly associated with disease^{1,2,3,4,5}. Changes in synonymous coding can alter mRNA splicing, mRNA folding, and stability^{6,7,8}, and can affect translational speed and accuracy and the conformation of the translated protein^{9,10,11,12}.
Numerous studies have shown that changes in the rhythm of translation can alter the kinetics of cotranslational folding and so, the global conformation of the final protein product^{13,14}. Translation rate is affected by synonymous codon usage which alters mRNA structure and tRNA abundance, the latter coevolving with codon bias^{15,16,17,18,19,20}. This mechanism provides an indirect association between codon usage and global protein structure. Nevertheless, whether and how synonymous variants of a gene will alter the conformation of the final folded protein is still poorly predictable and additionally the literature is riddled with reports of single synonymous mutations causing measurable functional effects that are not wellexplained by current mechanisms^{21,22,23,24}. Together this suggests that we are far from fully understanding the role of codon usage in orchestrating protein folding.
To the best of our knowledge, no studies have investigated whether the specific backbone torsion of an amino acid is associated with the synonymous codon from which it was translated. To probe for such a direct and local association, we developed a method for estimating and comparing codonspecific backbone dihedral angle distributions, which we term codonspecific Ramachandran plots. Comparing these distributions for pairs of synonymous codons, statistically significant differences are observed. Our results demonstrate that the backbone dihedral angle of an amino acid is statistically dependent on the identity of the codon from which it was translated, however, these results cannot shed any light on the causal direction of this dependence.
Results
Data collection, codon assignment and development of analysis tools
The first challenge in investigating the dependence between codon identity and the protein backbone structure is, regrettably, the absence of annotation within the Protein Data Bank (PDB) for the actual genetic template used in producing the protein for crystallization. Automatic assignment of codon identity to each position in a protein structure is a prerequisite to calculate codonspecific Ramachandran plots. It is imperative to stress that any method used for largescale codon reassignment will carry an inherent limitation of being contaminated with uncertainty and error. The main reason is that codon optimization is very commonly used to improve heterologous gene expression^{25}, especially in structure determination which necessitates the production of large amounts of soluble protein. There is not one common approach to codon optimization, and the choice of method often depends on trial and error^{26,27}. To limit codon assignment errors from including codonoptimized genes, we selected only structures of E. coli proteins expressed in E. coli, the most common expression system in the PDB. We purposely did not include all natively expressed proteins from other species, since codon biases differ between organisms^{28} and such generalization could obfuscate the sought for associations between coding and structure.
The procedure for computing and comparing codonspecific backbone dihedral angle distributions is displayed in Fig. 1 and detailed in the Methods. Briefly, highresolution PDB structures are retrieved, structures are filtered to remove homology bias and the resulting proteins are grouped according to their unique Uniprot entry. For each position in a protein chain, the backbone dihedral angles, φ and ψ, are calculated; if multiple PDB structures are available, the angles are averaged. Alongside this precise structural information, DSSP secondary structure is designated, and codons are assigned according to the genetic sequences obtained from ENA records crossreferenced in the Uniprot entry. Only locations with unambiguously assigned secondary structures and codons are retained.
We used only wellfitted Xray crystal structures having a resolution no worse than 1.8 Å (Supplementary Fig. 1), as a recent study considering alternate backbone conformations found resolutions better than 2.0 Å useful for such purposes^{29}.
A second challenge in investigating the dependence between codon identity and the protein backbone structure is that synonymous codons vary greatly in relative abundance (Supplementary Table 1). The challenge is that any difference we see in terms of the measured distance between estimated distributions, could be due to chance, arising from the availability of only finite data. Thus, our approach is to determine whether the distance we measure is large enough such that the probability of obtaining such a distance by chance from identical underlying distributions is extremely small.
Specifically, our analysis carefully accounts for this by comparing nonparametric distribution estimates which are calculated using the same sample size for all codons in a synonymous group; that of the rarest codon. We combine this with bootstrapresampling to account for all available data from the abundant codons. We do not assume any specific parametric form of the underlying (i.e., real) distribution of codon dihedral angles because these distributions are complex, unknown, and unlikely to be accurately approximated by any closedform parametric model. Instead, we aim only to compare the estimated distributions and as such do not require that samples from each codon distribution exhaustively represent the entire underlying distribution. Rather we developed tools and employed existing statistical methods which are sensitive and capable of comparing between nonparametric estimated distributions. We quantify the differences using a distributionfree statistical test to calculate the pvalue of observing these data under the assumption that the two codons in question have the same underlying distribution. The Methods section describes these details in full and shows additional experiments on synthetic and real data which validate our approach on various sample sizes.
Codonspecific backbone angle distributions are significantly distinct within the β mode
Synonymous codons are known to have distinct propensities to different secondary structures^{30,31,32,33}, which is manifested as different probabilities of the corresponding modes in the full codonspecific Ramachandran plots (Fig. 2). The difference in propensity for the main two, α and β, secondary structure modes might therefore dominate the difference between the codonspecific Ramachandran plots of synonymous codons. To factor out this effect, we conditioned the dihedral angle distribution on the secondary structure, effectively restricting it either to the distinct β or α modes. Select examples of the resulting codonspecific Ramachandran plots, conditioned on these modes, are shown (Fig. 3), while the full set is provided (Supplementary Figs. 2, 3). Visually, it is evident that synonymous codons of some amino acids have clearly distinguishable distribution shapes especially in the βmode.
To quantify those differences and their significance, we used a distributionfree twosample permutation test with the L_{1} distance between KDEs serving as the test statistic and assigned pvalues to each synonymous codon pair with respect to the null hypothesis that the two codons have the same underlying distribution of backbone angles. To determine the pvalue threshold for statistical significance in a setting where multiple hypotheses are considered together, we employed the BenjaminiHochberg correction with false discovery rate set to 0.05. This process is shown schematically (Fig. 1) and detailed in the Methods. Matrices and multidimensional scaling (MDS) plots visualizing the distances between select pairs of synonymous codon distributions are shown alongside the contour plots (Fig. 3) and for all synonymous codon groups (Supplementary Figs. 4–7).
Note that together with the 87 synonymous pairs, we also included the 61 comparisons of each codon to itself. The latter served as a control, and indeed, the null hypotheses were not rejected for any of the same codon pairs in either of the secondary structures. No synonymous pairs were rejected in comparisons of the distributions of the αmode, however, when comparing distributions for the β mode, 57 of the 87 synonymous pairs were rejected (Fig. 4).
It is not surprising that αhelices, being less flexible than βsheets^{34,35}, display less variability in codonspecific Ramachandran plots. The Ramachandran plot defines a richer range of structural contexts than the discrete categories available in DSSP annotation^{36}, especially in the βmode. It is therefore possible that some of the differences we observe between codonspecific dihedral angle distributions in the βmode are attributable to codon preferences for finer secondary structure categories such as parallel and antiparallel βsheets. However, it should be noted that we used a strict conditioning by the secondary structure, taking only the DSSP annotation^{37} E (extended strand – βsheet in parallel and/or antiparallel sheet conformation with minimum length of 2 residues) for the βmode, and H (αhelix – a 4 turn helix with minimum length of 4 residues), for the αmode.
Having found synonymous codons which have different dihedral angle distributions within the βmode, we explored the possibility that synonymous codon preferences for specific positions within this secondary structure (beginning, middle or end; refer to Supplementary Fig. 8)^{36} are reflected in these differences, at least in some amino acids. In Supplementary Fig. 9, we present codonspecific Ramachandran plots for secondary substructures of the βmode in an amino acid with large codon sample sizes (alanine). Substantial distribution differences are still observed between synonymous codons, even with such finer secondary structure conditioning. This indicates that distinct codon propensities for substructures of a βmode cannot fully explain their distribution differences observed in the full βmode.
Distances between dihedral angle distributions of synonymous codons hint at a correlation to features of the translation process
Our findings remain silent regarding the origin of the observed differences in synonymous codon backbone dihedral angle distributions; in particular, the causation direction cannot be established unambiguously. It is tempting to speculate, however, that the translation process plays an active role in the observed effect. To illustrate this speculation, we considered how two features of the translation machinery correlate to the calculated distances between backbone dihedral angle distributions of synonymous codon pairs. Firstly, we demonstrate that the difference in the codonspecific translation speed between a pair of synonymous codons appears to positively correlate to the distance between their dihedral angle distributions (Fig. 5, left). Although ribosome profiling has facilitated the measurement of translation speed to exquisite singlecodon resolution in human and yeast cells, the application to bacteria has been more problematic^{38}. We used the data from Chevance et al. who developed an in vivo bacterial genetic assay for measuring ribosomal speed independent of the stability of the mRNA transcript or the translated protein product^{39}. Note that in order to limit confounding factors, we considered only pairs of codons being translated by the same tRNA.
In a second illustration, we identified codons translated unambiguously by a single tRNA, following Bjork et al.^{40}, and grouped codon pairs as being translated by either the same or different tRNA molecules. Figure 5 (right) shows that synonymous codon pairs translated by different tRNAs tend to have a larger distance between their backbone dihedral angle distributions.
While these two trends can by no means be conclusive, they suggest the potential value of the proposed methods in analyzing relations between synonymous coding and the features of the translation process.
Discussion
In this work we generate codonspecific Ramachandran plots, showing that there is some association between synonymous codon usage and the structure of the translated amino acid. In contrast to previous works showing that synonymous codons have preferences for different secondary structures, this work probes for a much more local association, namely between the backbone dihedral angle distributions of an amino acid and the synonymous codon from which it was translated. To factor out the phenomenon of secondary structure preference, we analyzed the α and β modes separately and found that many synonymous codon distributions are statistically significantly different in the β mode. We found no statistically significant differences in distributions for the α mode; perhaps not surprisingly given that αhelices fold into more rigid structures. Although our results cannot determine causal direction, it is worth clarifying that should synonymous codon usage be found to directly affect the formation of local protein structure this would not challenge the dominance of the amino acid sequence and protein environment in directing protein folding, especially for globular proteins having a welldefined fold. We would suspect that only some positions in a structure could carry a memory of potential structural bias introduced by synonymous coding, and that any environmental effects will not be biased towards any particular codon. This would mean that although the inability to factor out the environment is a limitation of our study, the differences between synonymous codon distributions would underestimate the effect that synonymous coding could have at positions, which are sensitive to this effect.
Given the mounting evidence for an association between codon usage and protein structure, it is not surprising that there have been previous attempts to combine genetic information with the structural description of the protein products these genes produce. The Integrated SequenceStructure Database (ISSD) catalogued in its second edition 88 E. coli, 25 yeast and 166 mammalian nonhomologous proteins having a resolution better than 2.5Å^{41} and the CodConf Data Base assigned coding information to almost 1900 nonhomologous proteins from all species^{42}. Two important trends have eventuated in the two decades passed since the development of these databases: (1) there has been an exponential rise in the number of highresolution protein structures, and (2) codon optimization has become common place in heterologous gene expression for structural studies. This means that whilst we now have a wealth of structural data which could be used to explore associations with codons, they are not readily usable since structural databases, notably the PDB, rarely annotates the genetic sequence from which the protein was produced.
Codonspecific Ramachandran plots and their comparative analysis could serve as a useful, quantitative tool in future studies looking at the association between coding and local protein structure. It is likely that codonspecific backbone dihedral angle distributions will show even more significant variations when extended to pairs or triplets. Codon pair usage bias has been observed in E. coli^{43} and in human disease^{44,45}. It has been suggested that codon translation efficiency is modulated by adjacent single nucleotides^{46}, that codon pair order significantly affects translation speed^{40}, and, more recently, the case for a genetic code formed by codons triplets has been argued^{24}.
The main challenge in extending the presented methods to codon pairs or longer tuples is the relative scarceness of data and the need to compare multidimensional density functions characterizing the backbone structure of a tuple of amino acids. Extending the analysis to other expression systems faces a similar data scarceness challenge. Considering genes from various source organisms expressed in E. coli, either to probe for evolutionary distinctions between species or to overcome data scarceness when probing codon pairs in a hypothesized translationdependent mechanism, is burdened by the uncertainty associated with codon (re)assignment. The latter will be overcome when structural databases start annotating the exact genetic source used for producing protein, which is crucial, given the everamounting evidence for the critical functional importance of codon usage.
We hope that the associations between synonymous coding and local backbone conformation revealed through codonspecific Ramachandran plots will spark subsequent investigations which should directly probe the possible causal relationships that might underpin these observations. The implications of an active, codingdependent process would be tremendous, necessitating an immediate rethink as to how we manipulate the genetic code through codon optimization. This question could not be timelier, as mRNA vaccines are taking centre stage in global medicine. Moreover, the observed dependence between coding and local structure can potentially improve protein prediction algorithms, since in such tasks the causal relationship between the two is superfluous. To conclude, these results may affect how we define the role of synonymous variants in health and disease and understand protein folding in general.
Methods
Data collection
Protein structure data is collected from the Protein Data Bank (PDB)^{47} through a structured query against the search API defining the following criteria: (i) Method: XRay Diffraction; (ii) XRay Resolution: Less than or equal to 1.8 Å; (iii) R_{free}: Less than or equal to 0.24, (iv) Expression system contains the phrase “Escherichia Coli” and (v) Source organism taxonomy ID equal to 562 (Escherichia Coli). Queries return a list of PDB IDs having entity numbers, e.g., 1ABC:1. An entity corresponds to one or more identical polypeptide chains in the structure and we select the first of its matching chains using lexicographic order. A structure may have more than one unique entity (e.g., 1ABC:1 and 1ABC:2), in which case we would obtain both.
Next, we query the PDB’s entry data API to obtain a mapping from chains to Uniprot^{48} IDs. We keep only chains which map to a unique Uniprot ID, which is most chains. An exception, which we discard, are chimeric chains i.e. those that contain sections from multiple different proteins. We align the protein sequence of each chain to the Uniprot record sequence, using the same pairwise alignment algorithm as described in Codon assignment (below), to provide a Uniprot index for each residue in the PDB chain.
After removing homology bias using the procedure described under Redundancy filtering, backbone dihedral angles (φ,ψ) are calculated and secondary structure is assigned by DSSP^{37} per residue. Finally, we assign each residue with a codon using the method described below (codon assignment). The result of this process is what we call a Protein Record for each PDB chain. The Protein Record contains, per residue: corresponding Uniprot ID and residue index, torsion angles (φ,ψ), secondary structure and codon.
Redundancy filtering
To remove homology bias from our data, we performed a filtering step. First, each pair of Uniprot sequences is aligned using the BioPython^{49} software package, with a match score of 1 and all penalty scores set to zero. Thus, we obtain an alignment score s_{ij} ≥ 0 between every pair of Uniprot sequences i,j. We then calculate normalized scores,
Note that by definition \(0\le {\widetilde{s}}_{{ij}}\le 1\) and \({\widetilde{s}}_{{ii}}=1\) for every i,j. In other words, this normalization ensures that the selfalignment score is 1 and all other scores are normalized to be in [0,1], regardless of the sequence lengths or the alignment penalty values. This normalization also makes it simple to choose a similarity cutoff threshold, since the threshold is chosen in the fixed range [0,1] where 1 equates to an exact match and 0 to a complete mismatch. We chose a normalized similarity threshold of τ = 0.7.
Using the normalized alignment scores we then employ a farthestfirst traversal procedure to sort the Uniprot sequences: the first sequence is selected arbitrarily, and each successively selected sequence is such that it has the lowest maximum normalized alignment score between itself and all previouslyselected sequences. Formally, denote by \({{{{{\mathcal{S}}}}}}\) and \({{{{{\mathcal{U}}}}}}\) the sets of selected and unselected sequences, respectively. We initialize to \({{{{{\mathcal{S}}}}}}=\left\{0\right\}\) and \({{{{{\mathcal{U}}}}}}{{{{{\mathscr{=}}}}}}\left\{1,2,\ldots ,N1\right\}\) where N is the number of Uniprot sequences. At each step k of this traversal, for each unselected sequence \(j{{\in }}{{{{{\mathcal{U}}}}}}\), we calculate its greatest similarity to any of the sofar selected sequences,
We then choose the sequence which has the lowest maximal similarity to the selected sequences, i.e., we add
to \({{{{{\mathcal{S}}}}}}\). We stop the procedure once S_{k} (j) > τ for the sequence j that was selected at step k, and retain \({{{{{\mathcal{S}}}}}}\) as the output filtered set of Uniprot sequences. This ensures that no two sequences in the selected set have a normalized similarity score greater than τ. After performing this procedure, we keep in our dataset only PDB chains that were mapped to one of the Uniprot sequences in \({{{{{\mathcal{S}}}}}}\). Any PDB chain mapped to a Uniprot sequence from \({{{{{\mathcal{U}}}}}}\) is discarded from analysis. Note that we keep all chains from different PDB structures that correspond to the same selected Uniprot sequence in order to aggregate their backbone angles as explained below (under Angle aggregation).
Codon assignment
Since the genetic sequences used for expressing each protein are not annotated in the PDB, we assigned codons from the native sequence in the European Nucleotide Archive (ENA)^{50} IDs, crossreferenced from the mapped Uniprot ID. All available genetic sequences for the specific protein are translated to an aminoacid sequence and aligned pairwise to the sequence of the PDB chain. The alignment is performed using the BioPython^{47} implementation of the Gotoh global alignment algorithm^{51}. We used BLOSUM80 as the substitution matrix for the alignment, a gapopening penalty of −10 and a gap extension penalty of −0.5.
Following the pairwise alignment of the amino acid sequence to all translated genetic sequences, we obtain the aligned codons from each sequence and assign them to corresponding residues from the PDB chain. This process yields zero or more assigned codons per residue in the PDB chain. In cases where more than one codon is assigned, we consider the assignment ambiguous and exclude that residue from further analysis.
Angle aggregation
Since some proteins have been characterized by multiple crystal structures, there are residues from different PDB chains which map to the same Uniprot ID and location in the Uniprot sequence. For example, in our dataset, the residues 1SEH:A:42, 1RNJ:A:42 and 2HRM:A:42 were all aligned to the Uniprot ID and index P06968:41. We consider such cases as different experimental realizations of the same protein residue and aggregate the backbone angles from such residues, to obtain an average measurement.
When aggregating the angles, we must account for the fact that a torsion angle pair φ = (φ,ψ) is defined on a torus (i.e. the domain S^{1} × S^{1} where S^{1} is a circle). Intuitively, each angle naturally wraps around at ±180°, and the space spanned by two such angles is a torus. Thus, taking a simple average of each angle separately would not be correct. Instead we use the torusmean function defined in the mathematical tools section (below).
Backbone angle distribution distance
Our aim is to measure the distance between distributions of backbone torsion angles of synonymous codons in αhelix and βsheet secondary structure modes. Denote \(f\left({{{{{\mathbf{\varphi }}}}}}c,{{{{{\mathcal{X}}}}}}\right)\) the distribution of backbone angles φ of codon c in secondary structure \({{{{{\mathcal{X}}}}}}\). We denote the distance between the backbone angle distributions of two synonymous codons c and c′ in secondary structure \({{{{{\mathcal{X}}}}}}\) as \(d\left(c,c^{\prime} \right){{{{{\mathcal{X}}}}}}\) and estimate them between all pairs of synonymous codons. We include all cases of c = c′ as controls. Empirical tests with the L_{1,}, L_{2} and smoothed Wasserstein distances showed that the L_{1} metric provided the highest statistical power of all three at reasonable computational costs and so this was the distance metric selected. It is defined as,
Although the underlying backbone angle distributions \(f\left({{{{{\mathbf{\varphi }}}}}}{{{{{\rm{}}}}}}c,{{{{{\mathcal{X}}}}}}\right)\) are unknown, we sample from the distributions to obtain a finite sample \({\left\{{{{{{{\mathbf{\varphi }}}}}}}_{i}\sim f\left(\cdot {{{{{\rm{}}}}}}c,{{{{{\mathcal{X}}}}}}\right)\right\}}_{i}\) for each codon c and secondary structure \({{{{{\mathcal{X}}}}}}\). We use these samples to fit a kerneldensity estimate (KDE), \(\widehat{f}\left({{{{{\mathbf{\varphi }}}}}}{{{{{\rm{}}}}}}c,{{{{{\mathcal{X}}}}}}\right)\), of each distribution, as explained under Kernel density estimation (below). The distance metric \({d}_{1}\left(c,c^{\prime} \right)\vert{{{{{\mathcal{X}}}}}}\) is then calculated on the KDEs of each synonymous codon pair. Since the KDEs are discrete, the integration above becomes a sum,
where K is the number of KDE bins in each direction and \({{{{{{\mathbf{\varphi }}}}}}}_{{k}_{1},{k}_{2}}=\left({\varphi }_{{k}_{1}},{\psi }_{{k}_{2}}\right)\) are discrete evenlysampled grid points. We then use permutationbased hypothesis testing to determine whether the distance supports the (alternative) hypothesis that the codons have a significantly different distribution, as explained below.
Detecting synonymous codons with different angle distributions
Faced with finitesample estimations of codon backbone angle distributions, \(\widehat{f}\left({{{{{\mathbf{\varphi }}}}}}{{{{{\rm{}}}}}}c,{{{{{\mathcal{X}}}}}}\right)\), we aim to determine whether there exist pairs of synonymous codons \(\left(c,c^{\prime} \right)\) for which the underlying distributions, \(f\left({{{{{\mathbf{\varphi }}}}}}{{{{{\rm{}}}}}}c{{{{{\mathscr{,}}}}}}{{{{{\mathcal{X}}}}}}\right)\), are different. For every pair of synonymous codons and secondary structure \(\left(c,c^{\prime} \right){{{{{\mathcal{X}}}}}}\) (where we allow \(c=c^{\prime}\)), we define a null hypothesis, which states that they have identical underlying backbone angle distributions:
We used permutationbased hypothesis testing^{52} to obtain valid pvalues for each of these null hypotheses without the need to make assumptions about the backbone angle distributions \(f\left({{{{{\mathbf{\varphi }}}}}}{{{{{\rm{}}}}}}c{{{{{\mathscr{,}}}}}}{{{{{\mathcal{X}}}}}}\right)\) or the distribution of the distance metric \({d}_{1}\left(c,c^{\prime} \right){{{{{\mathcal{X}}}}}}\) under the null. The permutation testing procedure is detailed below (under Permutationbased twosample hypothesis test). We thus obtain, per secondary structure, a total of 148 pvalues: 61 for identical codons, \(c=c^{\prime}\), and an additional 87 for nonidential but synonymous codons, \(c\ne c^{\prime}\).
We used the BenjaminiHochberg method^{53} for multiple hypothesis testing. In this approach a significance threshold is calculated dynamically from the set of all obtained pvalues, in a way which controls the FalseDiscovery Rate (FDR) for the entire set of tests (instead of the typeI error of each individual test). The method allows us to specify the FDRcontrol parameter, q, and ensures that over repeated trials the expected value of the proportion between false discoveries (i.e. false rejections of the null hypotheses) and total discoveries (all rejections of null hypotheses) will be q.
Preventing bias due to sample size differences
One way to account for vastly different sample sizes this would be to crossvalidate the KDE kernel bandwidth and choose an appropriate value for each sample size. This is challenging, however, since we would need to separately crossvalidate for all codons, some with very limited data.
Instead, we opted to use a single kernel bandwidth, but fix the sample size for each set of synonymous comparisons. For each amino acid \({{{{{\mathcal{A}}}}}}\), and per secondary structure \({{{{{\mathcal{X}}}}}}\), we used the same, minimum, sample size \({N}_{{{{{{\mathcal{A}}}}}}{{{{{\mathscr{,}}}}}}{{{{{\mathcal{X}}}}}}}\) to estimate the distributions for all codons in a synonymous group. Due to computational constraints, we also set an upper limit N_{max} of 200. Thus, the sample size for all codons of amino acid \({{{{{\mathscr{A}}}}}}\) was calculated as
where \({N}_{c{{{{{\mathscr{,}}}}}}{{{{{\mathcal{X}}}}}}}\) is the sample size for codon c in secondary structure \({{{{{\mathcal{X}}}}}}\).
Having limited the sample size, we employed also a bootstrappedsampling^{54} scheme on top of the distribution estimation and comparison, so as to exploit all available data for more common codons. Specifically, for each codon \(c{{{{{\mathscr{\in }}}}}}{{{{{\mathcal{A}}}}}}\), we estimate its distribution B times from \({N}_{{{{{{\mathcal{A}}}}}},{{{{{\mathcal{X}}}}}}}\) samples drawn with replacement from its collected data. This gives us access to at most \(B\cdot {N}_{{{{{{\mathcal{A}}}}}},{{{{{\mathcal{X}}}}}}}\) samples from each codon \(c{{{{{\mathscr{\in }}}}}}{{{{{\mathcal{A}}}}}}\), instead of only \({N}_{{{{{{\mathcal{A}}}}}},{{{{{\mathcal{X}}}}}}}\). We then compare B pairs of distributions for each synonymous codon pair \(c,c^{{{\prime} }}{{\in }}{{{{{\mathcal{A}}}}}}\) using the permutation test, and use the results of all permutations in all bootstrap iterations to calculate the pvalue of \(\left(c,c^{\prime} \right)\).
Statistical tests were performed with B = 25 bootstrap iterations with K = 200 permutations each, for a total of 5000 permutations used for p value calculation. We used N_{max} = 200 for all comparisons and set an FDR threshold of q = 0.05. Figure 6 presents a syntheticdata experiment validating this approach using various sample sizes.
Full procedure
The procedure for comparing synonymous codon backbone angle distributions and then identifying codon pairs having significantly different distributions is described here.
For each synonymous codon pair, \(\left(c,c^{\prime} \right)\) and secondary structure \({{{{{\mathcal{X}}}}}}\), we calculate a pvalue with respect to the null hypothesis \({H}_{0,\left(c,c^{\prime} \right){{{{{\mathcal{X}}}}}}}\), i.e. that they come from the same underlying distribution:
For \(b\in \left\{1,\ldots ,B\right\}\):

Sample \({N}_{{{{{{\mathcal{A}}}}}}{{{{{\mathscr{,}}}}}}{{{{{\mathcal{X}}}}}}}\) observations randomly from c and from \(c^{\prime}\) (each with replacement).

Denote the sampled observations from c and \(c^{\prime}\) as \({{{{{\mathcal{C}}}}}}\) and \({{{{{\mathcal{C}}}}}}^{{{\prime} }}\) respectively.

Apply permutation test procedure (Permutationbased twosample hypothesis test) on \({{{{{\mathcal{C}}}}}}\) and \({{{{{\mathcal{C}}}}}}^{{{\prime} }}\) for K permutations. The teststatistic T(X,Y) first computes the KDEs of X and Y, then calcultates the L_{1} distance between them.

Denote by η_{b} the number of times the base metric no greater than the permuted metric in the current permutation test.
Calculate the pvalue with respect to \({H}_{0,\left(c,c^{\prime} \right){{{{{\mathcal{X}}}}}}}\):
For each secondary structure \({{{{{\mathcal{X}}}}}}\), we calculate the significance threshold based on the BenjaminiHochberg method as follows:

Denote \({\left\{{p}_{i{{{{{\mathscr{,}}}}}}{{{{{\mathcal{X}}}}}}}\right\}}_{i=1}^{M}\) the set of M = 148 pvalues obtained from all pairwise comparisons of synonymous codons in secondary structure \({{{{{\mathcal{X}}}}}}\).

Sort the pvalues and denote \({p}_{\left(i\right){{{{{\mathscr{,}}}}}}{{{{{\mathcal{X}}}}}}}\) the ith sorted pvalue.

Calculate the threshold pvalue index for an FDR of q, which is the largest p value smaller than the adaptive threshold of \(q\cdot i/M\):
$${i}_{0}={{{{{\rm{max }}}}}}\left\{i:{p}_{\left(i\right),{{{{{\mathcal{X}}}}}}}\le q\cdot \frac{i}{M}\right\}.$$ 
Set the adaptive significance threshold: \({\alpha }_{M}={p}_{\left({i}_{0}\right){{{{{\mathscr{,}}}}}}{{{{{\mathcal{X}}}}}}}\).

Reject the ith nullhypotheses if \({p}_{\left(i\right){{{{{\mathscr{,}}}}}}{{{{{\mathcal{X}}}}}}} < {\alpha }_{M}\).
The set of synonymous codon pairs corresponding to the rejected null hypotheses are deemed to have significantly different backbone angle distributions. Figure 7 presents an experiment on real codon data, visualizing how the pvalues calculated by our method correspond to expected differences between codons.
Mathematical Tools
Torus mean
Given a set of N points on a torus \({\left\{{{{{{{\mathbf{\varphi }}}}}}}_{i}\right\}}_{i=1}^{N}\) where \({{{{{{\mathbf{\varphi }}}}}}}_{i}=\left({\varphi }_{i},{\psi }_{i}\right)\in {S}^{1}\times {S}^{1}\), we would like to calculate the mean of these points, \(\bar{{{{{{\mathbf{\varphi }}}}}}}\) in a way which accounts for the wraparound of each angle at ±180°. We define a function which approximates a centroid on a torus, by calculating the average angle with circular wrapping in each direction separately. We denote this function as \(\bar{{{{{{\mathbf{\varphi }}}}}}}={{{{{\rm{torm}}}}}}\left(\left\{{{{{{{\mathbf{\varphi }}}}}}}_{i}\right\}\right)\). For example, if \({{{{{{\mathbf{\varphi }}}}}}}_{1}=\left(170,170\right)\) and \({{{{{{\mathbf{\varphi }}}}}}}_{2}=\left(170,130\right)\) then we expect \({{{{{\rm{torm}}}}}}\left(\left\{{{{{{{\mathbf{\varphi }}}}}}}_{1},{{{{{{\mathbf{\varphi }}}}}}}_{2}\right\}\right)=\left(180,160\right)\). We define the function as follows
where \({{{{{\rm{atan}}}}}}2\left(y,x\right)\) is a signed version of \({{{{{\rm{arctan }}}}}}\left(y/x\right)\) which uses the sign of both arguments to unambiguously recover the sign of the original angle θ such that y = sinθ and x = cosθ.
Torus distance
Given two points on the torus, \({{{{{{\mathbf{\varphi }}}}}}}_{1}=\left({\varphi }_{1},{\psi }_{1}\right)\) and \({{{{{{\mathbf{\varphi }}}}}}}_{2}=\left({\varphi }_{2},{\psi }_{2}\right)\), we measure the distance, in angles between these points using the torus distance function as follows:
Kernel density estimation
We used twodimensional kernel density estimation (KDE)^{55} to estimate backbone angle distributions from finite samples. Given samples \({\left\{{{{{{{\mathbf{\varphi }}}}}}}_{i}\right\}}_{i=1}^{N}\) from torsion angles of a codon c in secondary structure \({{{{{\mathcal{X}}}}}}\), we calculate
where φ represents points on a discrete grid, K is a scalar kernel function, tord(⋅,⋅) is the torus wraparound distance defined under Torus distance, and γ is a constant factor which normalizes the KDE so that it sums to one. The KDE was evaluated on a discrete grid of size 128 × 128, which corresponds to a bin width of \(360/128\approx {2.8}^{\circ }\). By applying the kernel to the wraparound distance, we correctly account for the distance on the torus between each sample and each grid point. We used a simple univariate Gaussian kernel, \(K\left(x\right)={{{{{\rm{exp }}}}}}\left({x}^{2}/2{\sigma }^{2}\right)\), with a variance of σ = 2 (equivalent to the kernel bandwidth). We used a fixed bandwidth for all KDEs, ensuring to always compare KDEs calculated from the same number of samples.
Permutationbased twosample hypothesis test
Given two statistical samples, \(X={\left\{{x}_{i}\right\}}_{i=1}^{{N}_{X}}\) and \(Y={\left\{{y}_{i}\right\}}_{i=1}^{{N}_{Y}}\) containing N_{X} and N_{Y} observations respectively, we wish to test whether the observations in both samples were obtained from the same underlying data distribution. A powerful and wellknown approach to do this, is by conducting a twosample statistical hypothesis test, with the null hypothesis that X and Y are sampled from the same distribution, i.e., \({H}_{0}:{P}_{X}\left(x\right)={P}_{Y}\left(y\right)\). Such a test allows one to determine whether there is sufficient evidence to reject the null hypothesis, while limiting the chance of a typeI error (false positive, or rejecting H_{0} when it is true) to be at most \(0 < \alpha \ll 1\). Denote by \(T\left(X,Y\right){\mathbb{\in }}{\mathbb{R}}\) a test statistic of our choosing, which numerically summarizes the differences between X and Y, such that the smaller the value of T(X,Y), the more X and Y are deemed similar. Further denote by \(\widehat{\tau }=T\left(X,Y\right)\) the value of this test statistic when evaluated on the samples at hand. To perform the hypothesis test, a pvalue is calculated, which is the probability of obtaining a result at least as large as \(\widehat{\tau }\) under the assumption that H_{0} is true: \(p={{{{{\rm{Pr }}}}}}\left[T\ge \widehat{\tau}  {H}_{0}\right]\). The null hypothesis H_{0} is then rejected if p < α, thereby limiting the probability of typeI error to be α.
In order to avoid making unfounded assumptions about the data or compromising on the choice of teststatistic, we employed a permutationbased twosample hypothesis test^{50}, where the distribution of TH_{0} can be estimated for any choice of T by randomly permuting the observations’ labels. The procedure can be described as follows:

Inputs: samples \(X={\left\{{x}_{i}\right\}}_{i=1}^{{N}_{X}}\), \(Y={\left\{{y}_{i}\right\}}_{i=1}^{{N}_{Y}}\), teststatistic \(T\left(X,Y\right){\mathbb{\in }}{\mathbb{R}}\), number of permutations K.

Compute the base statistic value: \(\widehat{\tau }=T\left(X,Y\right)\).

Pool the observations: \(Z=\left\{{z}_{1},\ldots ,{z}_{{N}_{X}+{N}_{Y}}\right\}=\left\{{x}_{1},\ldots ,{x}_{{N}_{X}},{y}_{1},\ldots ,{y}_{{N}_{Y}}\right\}\).

Compute a random permutation π of \(\left\{1,\ldots ,{N}_{X}+{N}_{Y}\right\}\), such that π(i) is the ith element of this permutation.

For \(k\in \left\{1,\ldots ,K\right\}\):

Permute the pooled observations: \({Z}^{\pi }=\left\{{z}_{\pi \left(1\right)},\ldots ,{z}_{\pi \left({N}_{X}+{N}_{Y}\right)}\right\}\).

Split the permuted observations:
$${X}^{\pi } = \left\{{z}_{\pi \left(1\right)},\ldots ,{z}_{\pi \left({N}_{X}\right)}\right\}\\ {Y}^{\pi } =\left\{{z}_{\pi \left({N}_{X}+1\right)},\ldots ,{z}_{\pi \left({N}_{X}+{N}_{Y}\right)}\right\}$$ 
Compute the permuted statistic value: \({\widetilde{\tau }}_{k}=T\left({X}^{\pi },{Y}^{\pi }\right)\).


Calculate \(\eta =\mathop{\sum }\limits_{k=1}^{K}{{{{{\bf{1}}}}}}\left[\widehat{\tau }\le {\widetilde{\tau }}_{k}\right]\), the number of times that the base statistical was no greater than the permuted statistic.

Calculate the pvalue \(p=\frac{1+\eta }{1+K}\).

Output: p and η.
The key observation behind this approach is that under the null, we can treat X and Y as labels which are randomly assigned to observations from the same data distribution. Therefore, by permuting the labels and calculating the permuted teststatistic, we are obtaining samples of TH_{0}. If H_{0} is indeed true, we expect that \(\widehat{\tau }\approx {\widetilde{\tau }}_{k}\), thereby yielding p ≈ 0.5 as K→∞. Conversely, if H_{0} is false, we would expect that \(\widehat{\tau } > {\widetilde{\tau }}_{k}\), and then p→0 as K→∞. In practice, the number of permutations K is limited by computational constraints. Nevertheless, since the smallest pvalue which can be obtained is \({p}_{{{{{{\rm{min }}}}}}}=1/\left(1+K\right)\), we know the upper limit for the number of necessary permutations for a given significance level (in case of a single test).
Pairwise distance plots
For each secondary structure \({{{{{\mathcal{X}}}}}}\) and each amino acid \({{{{{\mathcal{A}}}}}}\), the statistical test procedure outlined under Permutationbased twosample hypothesis test returns a matrix of all pairwise L_{1} distance statistics \(d\left(c,c^{\prime} \right){{{{{\mathcal{X}}}}}}\) averaged over bootstrap iterations, where \(c,c^{\prime}\) are two codons encoding \({{{{{\mathcal{A}}}}}}\). Since the used statistic is a metric, it is convenient to visualize it the form of a scatter plot in which each point represents a codon, and the Euclidean distances between each pair of points \(c,c^{\prime}\) approximate the distance statistic \(d\left(c,c^{\prime} \right){{{{{\mathcal{X}}}}}}\). Such scatter plots are typically produced using multidimensional scaling (MDS)^{56}. However, our case is different in the fact that the pairwise distance statistics are random variables, and the input data are finite sample approximations of their expected values. We devised a variant of multidimensional scaling capable of handling this setting, which as far as we know is novel.
We aim at finding a collection of isotropic twodimensional normal distributions \({{{{{\mathcal{N}}}}}}\left({\mu }_{c},{\sigma }_{c}^{2}I\right)\)with locations μ_{c} and scales σ_{c}, each representing a codon c. The scales represent the uncertainty in location and constitute an extension of the standard MDS procedure which considers only locations. A simple calculation shows that the difference between two samples randomly drawn from \({{{{{\mathcal{N}}}}}}\left({\mu }_{c},{\sigma }_{c}^{2}I\right)\) is itself normally distributed with zero mean and covariance \(2{\sigma }_{c}^{2}I\). Consequently, the squared Euclidean distance \({d}_{2}^{2}\left(c,c\right)\) is distributed as \(2{\sigma }_{c}^{2}\cdot {\chi }_{2}^{2}\), where \({\chi }_{2}^{2}\) denotes the chisquared distribution with two degrees of freedom. The expected value of the latter squared distance is given by 4\({\sigma }_{c}^{2}\) and should approximate the square of the measured statistic \({d}^{2}\left(c,c\right){{{{{\mathcal{X}}}}}}\) (the latter corresponds to the diagonal of the input distance matrix). We therefore determine the scale parameters by setting \({\sigma }_{c}=0.5 \cdot d\left(c,c\right){{{{{\mathcal{X}}}}}}\).
A similar reasoning applies to the offdiagonal entries: the squared Euclidean distance \({d}_{2}^{2}\left(c,c^{\prime} \right)\) is distributed as \({{}{\mu }_{c}{\mu }_{c^{\prime} }{}}_{2}^{2}+\left({\sigma }_{c}^{2}+{\sigma }_{c^{\prime} }^{2}\right)\cdot {\chi }_{2}^{2}\), and its expectation is therefore given by
and should approximate \({d}^{2}\left(c,c^{\prime} \right){{{{{\mathcal{X}}}}}}\). Defining the target pairwise distances
we now invoke a regular MDS to solve for the locations μ_{c}.
Note that the procedure is exact when the input statistics are Euclidean distances between twodimensional normal vectors; in other cases, the recovered locations and scales are merely an approximation of the underlying distributions.
For visualization completeness, we also report the averaged distance statistics. For convenience, the distances are normalized as \(\frac{d\left(c,c^{\prime} \right){{{{{\mathcal{X}}}}}}}{\sqrt{d\left(c,c\right){{{{{\mathcal{X}}}}}} \cdot d\left(c^{\prime} ,c^{\prime} \right){{{{{\mathcal{X}}}}}}}}\).
Dihedral angle distribution plots
The standard Ramachandran plot is often visualized either as a φ,ψ scatter of the individual samples or as a density image estimated using KDE (the latter is sometimes plotted as level contours). Often, regions containing a certain amount of probability are superimposed. However, none of these visualization techniques represent the amount of uncertainty in the finite sample estimate of the probability density function. To capture the latter, we devised a new visualization (Fig. 8), described below.
Given a sample \({\left\{{{{{{{\mathbf{\varphi }}}}}}}_{i}\right\}}_{i=1}^{N}\) of dihedral angles to visualize, we bootstrap B independent samples of size \({{{{{\rm{min }}}}}}\left\{N,{N}_{{{\max }}}\right\}\). A normalized density image \({f}_{b}\left({{{{{\mathbf{\varphi }}}}}}\right)\)is constructed from each sample \(b\in \{1,\ldots ,B\}\) using the KDE procedure outlined under Kernel density estimation. The density images are averaged into a single density image \(f\left({{{{{\mathbf{\varphi }}}}}}\right).\)
For the level contour \(\lambda \in ({{{{\mathrm{0,1}}}}})\), a threshold τ is calculated such that
To calculate the uncertainty region of the above contour, we calculate the threshold τ_{b} for each density image \({f}_{b}\left({{{{{\mathbf{\varphi }}}}}}\right)\) individually and produce a set of binary images containing 1 wherever \({f}_{b}\left({{{{{\mathbf{\varphi }}}}}}\right)\ge {\tau }_{b}\) and 0 elsewhere; such images represent the λsuper level sets of the \({f}_{{b}^{\prime}}\)s. We then average these binary images and calculate their α and (1 – α)level sets. The region between these two contours is shaded in the plot and represents the \([\alpha ,1\alpha ]\) confidence set.
In all our figures, unless specified otherwise, we used B = 1000 bootstraps with N_{max} = 200; three levels \(\lambda \in \{{{{{\mathrm{0.1,0.5,0.9}}}}}\}\) were plotted with confidence set to α = 0.1.
Reporting Summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.
Data availability
The protein records data collected for this study as well as full output datasets have been deposited in the Harvard Dataverse database [https://doi.org/10.7910/DVN/5P81D4].
Code availability
The code implementing the described data collection and analysis methods has been deposited in the Zenodo repository [https://doi.org/10.5281/zenodo.6345285]. Code is available under restricted access conditioned on user identification and agreement to academic use license.
References
Chen, R., Davydov, E. V., Sirota, M. & Butte, A. J. Nonsynonymous and synonymous coding SNPs show similar likelihood and effect size of human disease association. PLoS One 5, e13574 (2010).
Sharma, Y. et al. A pancancer analysis of synonymous mutations. Nat. Commun. 10, 2569 (2019).
Walsh, I., Bowman, M., Soto Santarriaga, I., Rodriguez, A. & Clark, P. Synonymous codon substitutions perturb cotranslational protein folding in vivo and impair cell fitness. Proc. Natl Acad. Sci. 117, 3528–3534 (2020).
Komar, A. The Ying and Yang of Codon Usage. Hum. Mol. Genet 25, R77–R85 (2016).
KimchiSarfaty, C. et al. A “silent” polymorphism in the MDR1 gene changes substrate specificity. Science 315, 525–528 (2007).
Mueller, W. F., Larsen, L. S., Garibaldi, A., Hatfield, G. W. & Hertel, K. J. The Silent Sway of Splicing by Synonymous Substitutions. J. Biol. Chem. 290, 27700–27711 (2015).
Pagani, F., Raponi, M. & Baralle, F. E. Synonymous mutations in CFTR exon 12 affect splicing and are not neutral in evolution. Proc. Natl Acad. Sci. 102, 6368–6372 (2005).
Zhou, X. et al. A Comprehensive Analysis and Splicing Characterization of Naturally Occurring Synonymous Variants in the ATP7B Gene. Front. Genet. 11, 592611 (2021).
Purvis, I. J. et al. The efficiency of folding of some proteins is increased by controlled rates of translation in vivo. A hypothesis. J. Mol. Biol. 193, 413–417 (1987).
Zhao, F., Yu, C. H. & Liu, Y. Codon usage regulates protein structure and function by affecting translation elongation speed in Drosophila cells. Nucleic acids Res. 45, 8484–8492 (2017).
Akashi, H. Synonymous codon usage in Drosophila melanogaster: natural selection and translational accuracy. Genetics 136, 927–935 (1994).
Drummond, D. A. & Wilke, C. O. Mistranslation induced protein misfolding as a dominant constraint on codingsequence evolution. Cell 134, 341–352 (2008).
Liu, Y. A code within the genetic code: codon usage regulates cotranslational protein folding. Cell Commun. Signal 18, 145 (2020).
Buhr, F. et al. Synonymous codons direct cotranslational folding toward different protein conformations. Mol. Cell. 61, 341–351 (2016).
Riba, A. et al. Protein synthesis rates and ribosome occupancies reveal determinants of translation elongation rates. Proc. Natl Acad. Sci. 116, 15023–15032 (2019).
Nackley, A. G. et al. Human catecholOmethyltransferase haplotypes modulate protein expression by altering mRNA secondary structure. Science 314, 1930–1933 (2006).
Bartoszewski, R. A. et al. A synonymous single nucleotide polymorphism in ΔF508 CFTR alters the secondary structure of the mRNA and the expression of the mutant protein. J. Biol. Chem. 285, 28741–28748 (2010).
Bulmer, M. Coevolution of codon usage and transfer RNA abundance. Nature 325, 728–730 (1987).
Ikemura, T. Correlation between the abundance of Escherichia coli transfer RNAs and the occurrence of the respective codons in its protein genes: a proposal for a synonymous codon choice that is optimal for the E. coli translational system. J. Mol. Biol. 151, 389–409 (1981).
Yulong, W., Silke, J. & Xia, X. An improved estimation of tRNA expression to better elucidate the coevolution between tRNA abundance and codon usage in bacteria. Sci. Rep. 9, 3184 (2019).
Karakostis, K. et al. A single synonymous mutation determines the phosphorylation and stability of the nascent protein. J. Mol. Cell Biol. 11, 187–199 (2019).
Rajeshbhai Patel, U., Sudhanshu, G. & Chatterji, D. Unraveling the Role of Silent Mutation in the ωSubunit of Escherichia coli RNA Polymerase: Structure Transition Inhibits Transcription. ACS Omega 4, 17714–17725 (2019).
Simhadri, V. L. et al. Single synonymous mutation in factor IX alters protein properties and underlies haemophilia B. J. Med Genet 54, 338–345 (2017).
Chevance, F. & Hughes, K. Case for the genetic code as a triplet of triplets. Proc. Natl Acad. Sci. USA 114, 4745–4750 (2017).
Angov, E., Hillier, C. J., Kincaid, R. L. & Lyon, J. A. Heterologous Protein Expression Is Enhanced by Harmonizing the Codon Usage Frequencies of the Target Gene with those of the Expression Host. PLoS ONE 3, e2189 (2008).
Fu, H. et al. Codon optimization with deep learning to enhance protein expression. Sci. Rep. 10, 17617 (2020).
Ranaghan, M. J., Li, J. J., Laprise, D. M. & Garvie, C. W. Assessing optimal: inequalities in codon optimization algorithms. BMC Biol. 19, 36 (2021).
Plotkin, J. B. & Kudla, G. Synonymous but not the same: the causes and consequences of codon bias. Nat. Rev. Genet. 12, 32–42 (2011).
Keedy, D. A., Fraser, J. S. & van den Bedem, H. Exposing Hidden Alternative Backbone Conformations in Xray Crystallography Using qFit. PLoS Comput Biol. 11, e1004507 (2015).
Adzhubei, A. A., Adzhubei, I. A., Krasheninnikov, I. A. & Neidle, S. Nonrandom usage of ‘degenerate’ codons is related to protein threedimensional structure. FEBS Lett. 399, 78–82 (1996).
Gu, W., Zhou, T., Ma, J., Sun, X. & Lu, Z. The relationship between synonymous codon usage and protein structure in Escherichia coli and Homo sapiens. Bio Syst. 73, 89–97 (2004).
Gupta, S. K., Majumdar, S., Bhattacharya, T. K. & Ghosh, T. C. Studies on the Relationships between the Synonymous Codon Usage and Protein Secondary Structural Units. Biochemical Biophysical Res. Commun. 269, 692–696 (2000).
Saunders, R. & Deane, C. M. Synonymous codon usage influences the local protein structure observed. Nucleic Acids Res 38, 6719–6728 (2010).
Emberly, E. G., Mukhopadhyay, R., Tang, C. & Wingreen, N. S. Flexibility of βsheets: Principal component analysis of database protein structures. Proteins: Struct., Funct., Bioinf 55, 91–98 (2004).
Emberly, E. G., Mukhopadhyay, R., Wingreen, N. S. & Tang, C. Flexibility of αhelices: Results of a statistical analysis of database protein structures. J. Mol. Biol. 327, 229–237 (2003).
Hollingsworth, S. A. & Karplus, P. A. A fresh look at the Ramachandran plot and the occurrence of standard structures in proteins. Biomolecular concepts 1, 271–283 (2010).
Kabsch, W. & Sander, C. Dictionary of protein secondary structure: pattern recognition of hydrogenbonded and geometrical features. Biopolymers 22, 2577–2637 (1983).
Mohammad, F., Green, R. & Buskirk, A. R. A systematicallyrevised ribosome profiling method for bacteria reveals pauses at singlecodon resolution. Elife 8, e42591 (2019).
Chevance, F. F., Le Guyon, S. & Hughes, K. T. The effects of codon context on in vivo translation speed. PLoS Genet 10, e1004392 (2014).
Björk G. R., & Hagervall T. G. Transfer RNA Modification: Presence, Synthesis, and Function. EcoSal Plus 6, (2014)
Adzhubei, I. & Adzhubei, A. ISSD Version 2.0: taxonomic range extended. Nucleic Acids Res. 27, 268–271 (1999).
Singh, V., Suri A. and ThomasCherian S. “CodConfDB: a codon  conformation database “ Proceedings of 2005 International Conference on Intelligent Sensing and Information Processing, 2005., pp. 355–358 (2005)
Yarus, M. & Folley, L. S. Sense codons are found in specific contexts. J. Mol. Biol. 182, 529–540 (1985).
Alexaki, A. et al. Codon and CodonPair Usage Tables (CoCoPUTs): Facilitating Genetic Variation Analyses and Recombinant Gene Design. J. Mol. Biol. 431, 2434–2441 (2019).
Diambra, A. Differential bicodon usage in lowly and highly abundant proteins. PeerJ., 5, e3081 (2017)
Cutler, R. W. & Chantawannakul, P. Synonymous codon usage bias dependent on local nucleotide context in the class Deinococci. J. Mol. Evol. 67, 301–314 (2008).
Sussman, J. L. et al. Protein Data Bank (PDB): Database of ThreeDimensional Structural Information of Biological Macromolecules. Acta Crystallogr. Sect. D: Biol. Crystallogr. 54, 1078–1084 (1998).
Apweiler, R. et al. UniProt: The Universal Protein Knowledgebase. Nucleic Acids Res. 32, D115–D119 (2004).
Cock, P. J. A. et al. Biopython: Freely Available Python Tools for Computational Molecular Biology and Bioinformatics. Bioinformatics 25, 1422–1423 (2009).
Leinonen, R. et al. The European Nucleotide Archive. Nucleic Acids Res. 39, D28–D31 (2010).
Gotoh, O. Optimal Sequence Alignment Allowing for Long Gaps. Bull. Math. Biol. 52, 359–373 (1990). 1990.
Chung, E. Y. & Romano, J. P. Exact and Asymptotically Robust Permutation Tests. Ann. Stat. 41, 484–507 (2013).
Benjamini, Y. & Hochberg, Y. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. J. R. Stat. Soc.: Ser. B (Methodol.) 57, 289–300 (1995).
Efron, B., and Tibshirani, R. J. An Introduction to the Bootstrap. CRC press (1994)
Simonoff, J. S. Smoothing Methods in Statistics. Springer Science & Business Media (2012)
Boyarski, A., and Bronstein, A. M. Multidimensional scaling. Computer Vision: A Reference Guide, Ikeuchi (Ed.) (2020)
Acknowledgements
We are grateful to Joel Sussman and John Moult for their constructive skepticism and valuable comments. We thank Yaniv Romano for his helpful discussions on statistical methods.
Author information
Authors and Affiliations
Contributions
AM posed the original hypothesis; A.R., A.M. and A.B. designed the studies, interpreted the results and wrote the manuscript; A.R. and A.B. developed all the computational methods and performed the analyses. A.R. and A.M. contributed equally to this work.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Peer review
Peer review information
Nature Communications thanks Henry van den Bedem and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Rosenberg, A.A., Marx, A. & Bronstein, A.M. Codonspecific Ramachandran plots show amino acid backbone conformation depends on identity of the translated codon. Nat Commun 13, 2815 (2022). https://doi.org/10.1038/s41467022303909
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41467022303909
This article is cited by

Variability in SOD1associated amyotrophic lateral sclerosis: geographic patterns, clinical heterogeneity, molecular alterations, and therapeutic implications
Translational Neurodegeneration (2024)

A dataset of alternately located segments in protein crystal structures
Scientific Data (2024)

Codon language embeddings provide strong signals for use in protein engineering
Nature Machine Intelligence (2024)

Machine learning classifiers predict key genomic and evolutionary traits across the kingdoms of life
Scientific Reports (2023)

Mechanisms and pathology of protein misfolding and aggregation
Nature Reviews Molecular Cell Biology (2023)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.