Proteomes differ among distant domains of lifeTwo competing hypotheses suggest that amino acid usage in proteomes differ or not among distant groups of living organisms. The lineage-specific hypothesis suggests that amino acid frequencies across proteomes are relatively constant even for distantly related species. On the other hand, the lifestyle hypothesis suggests that environment conditions, such as optimal growth temperature, are the primary drivers of amino acid usage differences. To test these, we first analysed the proteome of 5590 species with complete annotated genomes and identified taxonomy from the NCBI database (see Supplementary Materials for Materials and Methods and Extended Data 1). For each species, we calculated the amino acid profile as the sum of individual amino acid frequencies divided by the total sum of amino acid counts as in15, which resulted in amino acid profiles for 328 Archaea species, 4107 Bacteria species, 1118 Eukaryotes and 37 Viruses. In our data (F1,111792 = 7247.30, p < 0.001, Table S1), as in previous studies3, the number of redundant codons and the frequency for the corresponding amino acid were positively correlated and to account for this, we used standardised amino acid frequency (i.e. amino acid frequency divided by the number of redundant codons). We found no evidence that the amino acid frequencies were conserved across domains of life as shown by both differences in the principal component analysis (PCA) clusters (Fig. 1a, b) and in the average amino acid usage frequency across domains (Domain*Amino acid frequency: F3,111720 = 398.9, p < 0.001, Fig. 1c). This was confirmed using a PERMANOVA on the amino acid usage profile across domains of life (PERMANOVA: F3,5586: 328.27, p = 0.004). This effect was driven by proteome-wide differences in amino acid frequencies and not by a single or few amino acids having disproportionate effect, as shown in our PCA analysis (Fig. 1a) and the average amino acid usage frequencies across domains (Fig. 1c). Nonetheless, it is worth highlighting the nearly two-fold increase in frequency of cysteine (C) in eukaryotes and viruses compared with prokaryotes which is consistent with the literature25. The differences in amino acid profiles across proteomes were observed when we analysed proteomes without codon standardisation (Supplementary file 1; F3,111720 = 419.57, p < 0.001) as well as for proteomes standardised using amino acid molecular weight, which is known to correlate negatively with amino acid frequency24 (Supplementary file 1; Domain*Amino acid: F3,111720 = 452.12, p < 0.001). These results show that amino acid profiles across proteomes are not conserved in species from distant domains of life.Fig. 1Amino acid frequencies across proteomes and the effects of growth temperature. (a) Principal component analysis (PCA) on standardised amino acid frequencies reveals that amino acid usage across genomes in the four domains of life differ. Standardisation was done by dividing the amino acid frequency by the number of redundant codons (see “Materials and methods”). Dots represent the centroids for each of the clusters (domains of life). (b) Sampling distribution of Hausdorff distances comparing the PCA clusters. Values for the Hausdorff distances equating zero represent no differences between two sets of points (see “Materials and methods” for details). (c) Average standardised amino acid frequencies in proteomes across four domains of life. (d) The relationship between standardised amino acid frequencies and optimal growth temperature reveals that environment only has a minor effect on amino acid profiles. In panel (d), axes were ln-transformed. Note that (a–c) tests the lineage-specific hypothesis whereas (d) tests the lifestyle hypothesis (see Main Text).Environment weakly explains variation in proteomesPast studies hypothesised that variation in amino acid profiles were driven by environmental conditions such as optimum growth temperature5,14,15,26 (the lifestyle hypothesis). Thermophilic proteomes were shown to be under stringent evolutionary constraints27 which affect amino acid profiles relative to mesophilic proteomes4,26 and can lead to highly contrasting amino acid usage profiles4. However, these studies were taxonomically biased because they lacked representation of large sets of mesophilic archaea4,14,26. We incorporated to our proteome dataset information on optimal growth temperature (in oC) for 296 species of Eukaryotes (n = 5), Bacteria (n = 149) and Achaea (n = 141) obtained from the ThermoBase database28 and the supplementary data in24 (Extended Data 2). From those, 65.8% (n = 195) were mesophilic (optimal growth up to 70 °C) whereas the remaining 34.2% (n = 101) were thermophilic (optimal growth between 70 and 110 °C). We firstly measured how much additional variance was explained when temperature was added as covariate in the model of amino acid frequencies for the 296 species for which growth temperature was available. Temperature had a statistically significant but small contribution to explaining the variance in our model (Likelihood ratio: \(\chi\)2-value: 1113.2, p < 0.001; R2 with vs without growth temperature: 0.831 vs 0.797), suggesting that the effects of growth temperature on amino acid profiles were small. One explanation for this is that only few amino acids which are thermolabile respond negatively to growth temperature and thus, the potential for variation at the proteome level is masked by other amino acids. These results do not contradict previous comparisons of proteomes using pairwise or sophisticated data transformations4,14,26 but they show that the magnitude of the influence of the environment on the proteome is minor.Next, we investigated the strength of the relationship between amino acids and growth temperature to disentangle which amino acids, if any, were linked to increasing growth temperatures. For instance, cysteine has been considered an environment-sensing amino acid and is also thermolabile29. Our data showed that amino acid frequencies differed with increasing growth temperature among domains of life (Growth Temperature*Domain*Amino acid: F38,5800 = 3.005, p < 0.001; Fig. 1d). This was driven primarily by the an increase in frequency of Alanine (A) and Arginine (R) alongside a strong decrease in frequency of Asparagine (N) and Lysine (K) with increasing growth temperature in eukaryotes but not in archaea or bacteria, a decrease in frequency of Aspartic acid (D) with increasing growth temperature in archaea but not in bacteria and eukaryotes, and an increase in the relative frequency of Glutamic acid (E) and Phenylalanine (F) with increasing growth temperature in bacteria but not archaea or eukaryotes (Fig. 1d). There were also statistically significant main effects (Amino acid: F19,5800 = 5.526, p < 0.001, Table S1) and two-way interactions (Growth temperature*Amino acid: F19,5800 = 2.384, p < 0.001; Domain*Amino acid: F38,5800 = 3.280, p < 0.001, Table S1) on amino acid frequency. These results confirmed a previous report26 that increasing growth temperature leads to an overall decrease in frequency of the thermolabile amino acids such as cysteine (C) and glutamine (Q)5,30,31, a pattern which we observed in our data in eukaryotes, archaea and bacteria. Cysteine has also been considered an anomalous amino acid due to lower frequencies than expected by cost models across proteomes14,24,30 although these studies did not directly control for growth temperature which could explain the relatively lower cysteine frequency than expected. Nevertheless, our results show that despite its low frequency, thermolabile amino acids in the proteome negatively correlate with higher optimal growth temperatures. More broadly, our results show that environmental effects in proteomes are minor.Few amino acids predominantly appear in first and last ranksWe then ranked amino acids from most to least frequently used in proteomes to investigate their usage frequencies by rank. Our data shows that the proportion of amino acid by rank were similar across domains of life (Domain*Rank*Amino acid: F57,771 = 1.097, p = 0.294), supporting the assumption that amino acid usage by rank is shaped by cost-minimization constrains across domains of life. Our data also showed that amino acids were not used uniformly across ranks (Rank*Amino acid: F19,888 = 9.376, p < 0.001; Fig. 2a). This suggested that some amino acids might be differentially prevalent (or even altogether absent) across ranks, which could highlight patterns of amino acid preferential use or avoidance.Fig. 2Amino acid diversity decreases in high and low usage ranks. (a) Amino acid proportions by rank across domains of life. (b) Amino acid diversity calculated as the Shannon–Wiener index (see “Materials and methods”) by rank, showing that diversity decreases at the higher and lower ranks. This means that only few amino acids are frequently or rarely used, while almost all amino acids can be seen at intermediate ranks.Although we did not measure amino acid usage costs directly, our rationale was that amino acid usage by rank could reflect the costs of amino acid usage assuming cost-minimization32. In this context, we predicted that (a) few amino acids are physiologically cheap and/or have been incorporated first into the genome to be highly abundant, (b) many amino acids have intermediate frequencies and (c) few amino acids are physiologically expensive and/or have been newly incorporated into genomes, thus having relatively low frequencies. If this pattern is consistent across taxa and domains of life, this would generate an inverted U-shape curve when we measure the diversity of amino acids that are used relative to their rank frequencies, with few amino acids in the most and least frequent ranks and many amino acids at intermediate ranks. Thus, under cost-minimization, only few amino acids are expected to be most or least frequently used, leading to a non-linear relationship between amino acid diversity and frequency ranks. This lower diversity at the edges of the amino acid frequency ranks resemble similar edge effects found in ecology33,34. We therefore tested whether amino acid usage by rank displayed such edge effect in proteomes across domains of life. To test this, we analysed the diversity of amino acids within each rank to assess how many amino acids (raw counts) and their weighted proportions (Shannon–Wiener diversity index) were present across ranks (see Eq. 1 in “Materials and methods”). Our data showed that amino acid diversity by rank varied linearly and non-linearly with rank across domains in both raw counts (Rank*Domain: F3,68 = 3.962, p = 0.011; Rank2*Domain: F3,68 = 3.408, p = 0.022, Table S1) and Shannon diversity index (Rank*Domain: F3,68 = 3.092, p = 0.032; Rank2*Domain: F3,68 = 6.438, p < 0.001). It also confirmed the strong non-linearity of amino acid usage by rank (Counts: Rank2: F1,68 = 578.117, p < 0.001; Shannon: Rank2: F1,68 = 425.21, p < 0.001, Table S1) which was not observed for the linear term (Counts: Rank: F1,68 = 0.001, p = 0.967; Shannon: Rank: F1,68 = 0.987, p = 0.323). These results corroborate our predictions and highlight a novel edge effect where the diversity of amino acids that appeared in ranks 1–2 and 19–20 was lower than the diversity of amino acids in intermediate ranks, an effect observed across all domains of life (Fig. 2b). It is unlikely that the edge effect was a statistical artifact because it was observed when the data was analysed without codon standardization or with standardization by amino acid molecular weight (Supplementary File 1) and for proteomes from increasing growth environment (Supplementary File 2).The edge effect is present in the amino acid profiles of secondary structuresThe edge effect appears to be ubiquitous and thus, we hypothesised that its cause must also be rooted into fundamental biophysical principles shaping amino acid usage. It is well established that secondary structures such as \(\alpha\)-helices and \(\beta\)-strands are often conserved among protein superfamilies even in distantly related species 35,36,37,38. Moreover, amino acids differ in their propensity to form \(\alpha\)-helices and \(\beta\)-strands39,40 which could influence how often they are used in proteomes, depending on their role in secondary structures. Thus, we hypothesised that amino acid frequencies in the proteome reflected their propensity to appear in protein secondary structures, such as \(\alpha\)-helices and \(\beta\)-strands, which could explain the edge effect and why few amino acids appear in most and least frequent ranks in proteomes in all domains of life. To test this, we analysed the amino acid frequency in secondary structures of 40,885 PDB unique entries from 3512 species across all four domains of life, selected from a subset of structures with low sequence similarity and solved at high-resolution (< 3 Å) from the PISCES culling database41,42. We first tested whether the average amino acid frequency in the proteome correlated with the average frequency of the amino acid in \(\alpha\)-helices and \(\beta\)-sheets and found that the average frequency in the proteome and secondary structures were statistically correlated in \(\alpha\)-helices (Frequency SSE: F1,72 = 92.08, p < 0.001) and \(\beta\)-sheets (Frequency SSE: F1,72 = 8.846, p = 0.003) but differed across domains of life for both secondary structure types (Domain*Frequency SSE \(\alpha\): F3,72 = 28.279, p < 0.001; Domain*Frequency SSE \(\beta\): F3,72 = 11.870, p < 0.001, Table S1). This was driven by a weaker positive relationship between amino acid frequencies in the proteome and \(\alpha\)-helices and a negative relationship between amino acid frequencies in the proteome and \(\beta\)-sheets in Viruses compared with other domains of life (Fig. 3a).Fig. 3Edge effect is likely driven by conformational properties of amino acids. (a) The relationship between average standardised amino acid frequency in the proteome (y-axis) and on the secondary structures (x-axis). (b) Amino acid diversity by rank, calculated using the Shannon–Wiener index, within secondary structures. (c) Amino acid diversity by rank, calculated using the Shannon–Wiener index, of simulated proteins with varying mixtures of α-helices/β-strands ratio. Note that medium and long SSE length tend to overlap. (d) Comparison between amino acid diversity by rank calculated using the Shannon–Wiener index from the proteome analysis (Observed) and from simulated data using propensity to form secondary structure from37 [Simulated (Propensity)].Next, we hypothesised that the proteome-level edge effect could be an emerging property of the edge effects in secondary structures. We tested this by first ranking amino acids from most to least frequently used in either \(\alpha\)-helices or \(\beta\)-strands for each species and measured the diversity of amino acids in each rank across the domains of life. There was evidence of a non-linear edge effect in both \(\alpha\)-helices and \(\beta\)-strands (Rank2: F1,144 = 184.833, p < 0.001) but this varied depending on secondary structure and domain. For instance, the edge effects on \(\beta\)-strands were relatively stronger in higher ranks as opposed to lower ranks, while the edge effects on \(\alpha\)-helices were more relatively symmetric showing a more characteristic inverted U-shape curve (Rank2*SSE: F1,144 = 21.719, p < 0.001; Fig. 3b). The non-linearity pattern of the rank-frequency curve for \(\beta\)-strands was less accentuated in viruses which drove a weak but statistically significant interaction between domain and the non-linear effect of rank (Rank2*Domain: F3,144 = 2.779, p = 0.043; Fig. 3b, Table S1). These results show that the edge effect observed at the proteome-level was also present in the frequency rank of secondary structures, suggesting that the edge effect at the proteome level could be an emerging property of how amino acids form secondary structures.The edge effect on amino acid diversity emerges from amino acid conformational propertiesOur results for amino acid diversity in secondary structures resembled the distribution of amino acid propensities to form \(\alpha\)-helices or \(\beta\)-strands reported in the classical work by Chou and Fasman39, where distributions of amino acid propensities for \(\alpha\)-helices were relatively symmetric while the propensity for \(\beta\)-strands were rightly skewed (Supplementary File 3). This led us to hypothesise that amino acid secondary structure propensities could be the biophysical constrain that gives rise to the edge effect at the proteome level, because amino acids could be differentially selected and used based on their secondary structure propensities. To test whether secondary structure propensity could alone replicate our findings, we simulated 54,400 sequencies of \(\alpha\)-helices and \(\beta\)-strands of varying lengths (from 6 to 66 in increments of 8 residues) where amino acids composition of these simulated secondary structures were selected with probability based only on their propensity to form \(\alpha\)-helices and \(\beta\)-strands as in39. This gave us a pool of \(\alpha\)-helices and \(\beta\)-strands with varying amino acid profiles which were representative of their secondary structure propensities. From this pool of simulated secondary structures, we randomly sampled \(\alpha\)-helices and \(\beta\)-strands to assemble 153,450 virtual proteins containing a mixture of these secondary structures. We simulated virtual proteins that were small (2–20 secondary structures), medium (21–40 secondary structures) or large (41–66 secondary structures), each of these with a mixture of secondary structures ranging from proteins that were primarily made of \(\alpha\)-helices (90–10%), balanced (50–50%) or \(\beta\)-strands (90–10%). As expected, the mixture of secondary structures (F4,8970 = 178.066, p < 0.001) and length (F2,8970 = 162.58, p < 0.001) of the simulated proteins influenced amino acid rank diversity. However, there was strong evidence that the edge effect could be rescued (Rank2: F1,8970 = 258.30, p < 0.001; Fig. 3c) independently of length and mixture (Length* Rank2: F2,8970 = 0.069, p = 0.932; Mixture* Rank2: F4,8970 = 0.790, p = 0.531; Fig. 3c, Table S1). The edge effect disappeared in simulations where amino acids were drawn with equal probabilities (Supplementary File 4 and Supplementary File 5), supporting that the edge effect is an emerging property of amino acid-specific secondary structure propensities.We then tested how the simulations compared to our observed data in relation to the edge effect. To do this, we compared the rank-frequency curve from the proteomes in our data base with the curve obtained from the simulations using propensity to form secondary structures from literature39. We recapitulated the same edge effect observed in our proteome dataset with our simulation parameterised solely with amino acid secondary structure propensity as in 39. Both amino acid diversity (Rank2*Data type (simulated vs observed): F1,36 = 0.172, p = 0.680) and amino acid count showed evidence of edge effects comparable to our observed data (Rank2*Data type (simulated vs observed): F1,54 = 0.454, p = 0.504; Fig. 3d). Simulations had lower raw amino acid counts per rank (F1,36 = 28.963, p < 0.001, Table S1) although not lower diversity (F1,36 = 3.135, p = 0.085; Fig. 3d). These results confirm that amino acid secondary structure propensities could underpin the edge effect on amino acid rank diversity.Amino acid usage rank is independent of their evolutionary originThe evolutionary origin of amino acids into the proteome could influence the frequency in which they are used and therefore, contribute to the edge effect. Two consensus sequencies of amino acid evolutionary chronology exists43 and we tested whether the average rank of amino acids based on their usage matched their average rank based on their evolutionary chronology. There was no evidence that average amino acid rank from frequency correlated with average rank from evolutionary chronology across domains of life for neither (Raw: F3,72 = 0.405, p = 0.749; Filtered: F3,72 = 0.390, p = 0.759; Fig. 4, Table S1). These results show that amino acid rank usage is determined by functional constrains on their use above evolutionary chronology. It is possible that the relationship between amino acid evolutionary origin in the genetic code and average rank is present for ancestral coding genes but not for genes that evolved more recently44. This could mask the relationship between evolutionary chronology and average amino acid rank computed here. Future studies which incorporate the evolutionary history of individual genes will help elucidate this. Nonetheless, our results suggest that the edge effect at the proteome level is unlikely to be driven by asymmetries in amino acid evolutionary chronology.Fig. 4Relationship between average rank from amino acid usage and average rank from evolutionary chronology as in Trifonov43. (a) ‘Raw order’ means unfiltered average chronology ranks from the 40 criteria and (b) ‘Filtered order’ means the filtered average chronology ranks, accounting for correlation between the 40 selection criteria as in Trifonov43. 95% confidence intervals on the slopes shows that none of the slopes are statistically significant (Table S1).
