Protein sequences were downloaded from Uniprot Knowledgebase (UniprotKB), a freely available online database of protein sequences and functional information (Apweiler et al., 2004; Bateman et al., 2023). The amino acid sequences were generally derived from genetic sequencing studies. UniprotKB contains almost 250 million protein sequences, including approximately 181,056 sequences corresponding to the RuBisCO large subunit (rbcL gene) and 4,491 sequences corresponding to the small subunit (rbcS gene) (April 2025). The majority of RuBisCO protein sequences in UniprotKB are from land plants. UniprotKB is composed of two sets of data, UniprotKB/Swiss-Prot and UniprotKB/TrEMBL. Data in UniprotKB/TrEMBL are unreviewed protein sequences, with computationally generated annotation and functional characterization. UniprotKB/Swiss-Prot consists of a high quality, manually curated and annotated, non-redundant protein sequence database. Data from both the Swiss-Prot and TrEMBL components of UniprotKB were used in this study. Most sequences were from TrEMBL, as Swiss-Prot represents < 0.25% of sequences. The protein sequences analyzed were representative of precursor proteins, i.e. proteins before post-translational modification.

The sequences were downloaded in Microsoft Excel format (.xlsx) during July 2023. Separate files were downloaded for each taxonomic group of photosynthetic organisms and each of the two genes in Form I RuBisCO. As the focus of this work was the major groups of photosynthetic organisms extant in the ocean, the taxonomic focus was families of aquatic photosynthetic eukaryotes and the cyanobacteria (Table 1). For comparison, data for dicotyledon flowering plants (class Magnoliopsida) were also used (Table 1). Only the Swiss-Prot subset of UniprotKB data were used for the analysis of flowering plants to reduce biases caused by the large size of this dataset compared with the others. Secondly, only using Swiss-Prot reduced the bias and redundancy of multiple sequences from a relatively small group of crop plants and model species (e.g. Arabidopsis thaliana).

The data were processed to eliminate incomplete sequences or sequences annotated as having potentially significant structural issues. Most of the amino acid sequences representing the large and small subunits were annotated as ‘fragments’ and therefore did not represent the full length of the protein. All fragments were discarded and not used in further analysis. A conservative approach was used to eliminate sequences within the remaining pool containing possible sequence errors. Sequences flagged with ‘sequence caution’, ‘sequence conflict’ or ‘caution’ labels were eliminated from the data, with each of these categories of warning indicating a range of different potential issues with the sequence. A very few sequences contained unknown or atypical amino acids within the sequence, indicated by the presence of ‘X’. These sequences were not analyzed as subsequent calculations assume that each amino acid in the sequence can be identified.

A significant number of the remaining sequences were likely to be incomplete or fragments, however they were not identified by the automated annotation within UniprotKB. In some cases, these sequences were less than half or more than double the length of reviewed sequences for that group of organisms, indicating that they were highly unlikely to represent a functional subunit. The elimination of short and long sequences were performed consistently, allowing for variation in sequence length (N) without biasing the data by adding erroneously short or long protein sequences. The ‘normal’ sequence length for each protein subunit for each taxonomic group was defined as the mode amino sequence length. In most cases, the mode sequence length corresponded to the length of reviewed reference sequences (Swiss-Prot) for that taxonomic group. Sequences outside of a range determined by the mode sequence length (N_mode) ± 2% were eliminated from the data. For example, the mode sequence length of the large subunit (encoded in the rbcL gene) in diatoms (Bacillariophyta) is 490 amino acids, therefore sequences ranging from 490 ± 10 amino acids (rounding to the nearest whole amino acid) were included in the analysis (Table 1). For diatoms, large subunit sequences outside of the range 480 to 500 amino acids were eliminated from further analysis. Finally, replicated sequences from the same operational taxonomic unit were removed. Replicates were defined as coming from the same species of eukaryote or the same strain of cyanobacterium. Strains of cyanobacteria, rather than species, were considered more appropriate as there is a significant genetic diversity within ‘species’ of cyanobacteria such as Prochlorococcus marinus (Biller et al., 2015). In the few cases where replicates were present in the data, the first sequence in the list was retained and the subsequent replicates were deleted.

The content of each element in a protein subunit was determined based on the total number of each specific amino acid in the sequence and the elemental content of each amino acid (Supplementary Table S1). This approach was used to calculate the number of C, N and S atoms in each protein subunit analyzed. The content of H and O in each protein subunit was determined using the same approach, with the addition of a correction accounting for the loss of a water molecule (H₂O) through the formation of peptide bonds between two amino acids. The number of peptide bonds in a protein subunit is N-1, resulting in the loss of N-1 oxygen atoms and 2(N-1) hydrogen atoms.

A similar approach was used to estimate the amount of energy required to synthesize each protein subunit. These determinations are not absolute, but useful for comparing the relative amount of energy required to assemble RuBisCO in different organisms. The energetic cost of building a protein can be divided into the direct costs of assembling the protein and indirect energetic costs associated with supporting the process. Only direct costs were accounted for in this calculation, specifically the energetic cost of building each amino acid in the protein, plus the energetic cost of assembling those amino acids into proteins. The energy used to synthesize proteins can be quantified in terms of high-energy phosphate bonds (~P) and reducing power (H) (Akashi and Gojobori, 2002). These costs are associated with the requirement of adenosine triphosphate (ATP) and Guanosine-5’-triphosphate (GTP) and hydrogen atoms (from nicotinamide adenine dinucleotide (NADH), nicotinamide adenine dinucleotide phosphate (NADPH), and dihydroflavine adenine dinucleotide (FADH₂) molecules). High energy phosphate bonds (~P) were used as the energy unit in these calculations, assuming 2 phosphate bonds are required per H (Akashi and Gojobori, 2002). The energetic costs of synthesizing each of the 20 amino acids found in proteins was from Akashi and Gojobori (2002) (Supplementary Table S1). These energetic costs were determined for the bacterium Escherichia coli, therefore it is assumed that these values are representative for the diverse range of bacteria and eukaryotes investigated in this work. This assumption was checked by comparison with Wagner (2005), who calculated the energetic cost of synthesizing amino acids in the yeast Saccharomyces cerevisiae (a eukaryote) and found similar values to Akashi and Gojobori (2002) (Supplementary Table S1). While the values were not identical, there was a significant positive correlation (r = 0.950; n = 19) between the energetic costs for the 20 amino acids determined in the two studies (Supplementary Figure S1).

The energetic cost of synthesizing the amino acids in each protein subunit were determined by multiplying the count of each amino acid in the sequence by the energetic cost of synthesizing that amino acid. These values were added together for each of the 20 amino acids to calculate a total energetic cost of synthesizing all the amino acids in the protein subunit. A cost of 4.5 to 5.9 ~P per amino acid is the estimated energetic cost for polymerizing amino acids into a primary protein structure (Amthor, 2000). The median value of this range (5.2 ~P per amino acid) was used to account for this energetic cost in this study, therefore the cost of assembling all the amino acids into the protein subunit was 5.2(N-1). The total energetic cost for synthesizing the primary structure of the protein subunit was calculated by adding together the energetic costs associated with synthesizing the all the amino acids in the protein sequence and the cost of assembling them into a protein.

The total cost of synthesizing the hexadecameric holoenzyme was calculated by simply adding up the costs of synthesizing the 8 large subunits and 8 small subunits to form the final L₈S₈ structure. Unaccounted energetic costs include those associated with the correct folding of the subunits, and transport to the site of assembly. The final assembly of L and S subunits into the L₈S₈ structure would incur energetic costs from binding and associated chaperone proteins and cofactors required to assemble the final molecule (Aigner et al., 2017; Yeates and Wheatley, 2017). These costs were not included in the estimate as they are difficult to determine and are a step removed from the basic research question. Therefore, the energetic costs calculated here represent the cost of producing the primary structures of the 16 protein subunits required to build a Form I RuBisCO molecule.

Data were plotted and analyzed using SigmaPlot 15.0 (Grafiti LLC.). Analysis of variance (ANOVA) was conducted on data that met the assumptions of normality and equality of variance. The Holm-Sidak method was used to make post-hoc pairwise comparisons. The Kruskall-Wallace one-way analysis of variance on ranks (H) was used on data did not meet these assumptions. Post-hoc pairwise comparisons were conducted using Dunn’s Method. Data were pooled to compare the properties of RuBisCO proteins between eukaryotes and bacteria. The non-parametric Mann-Whitney U test (U) was used to compare the median properties of RuBisCO proteins between eukaryotes and bacteria as these pooled data violated the assumption of normality and therefore a parametric t-test was not suitable.

The taxonomic groups that were included in the analysis, and which type of RuBisCO they contain, are listed in Table 1. Phytoplankton are usually categorized in terms of phylogenetic groups, but types of RuBisCO are not associated with single clades of organisms due to the endosymbiotic origins of different groups. Consequently, the data were summarized using a scheme that emphasizes both taxa and type of RuBisCO. Cyanobacteria were divided into two groups, associated with Type IAc and Type IBc RuBisCO. The eukaryotes were grouped by supergroup (according to Keeling and Burki, 2019). However, the Archaeplastida were split into two groups according to whether the included taxa were ‘red’ (Rhodophyta) or ‘green’ (Viridiplantae), containing Types ID and IB RuBisCO, respectively (Leebens-Mack et al., 2019).

Despite the large number of RuBisCO sequences in UniprotKB, data filtering resulted in a total of only 41 complete holoenzyme sequences (Table 2). RuBisCO molecules in cyanobacteria (both Types 1Ac and 1Bc) contained > 390 less amino acids than the eukaryotes. There was a positive correlation between the number of carbon and nitrogen atoms in the L₈S₈ holoenzyme of RuBisCO (Figure 2A), with different groups of organisms clustering based on the type of RuBisCO. Cyanobacteria (Types 1Ac and 1Bc) contained significantly (p < 0.05) less C and N than the Viridiplantae (Type 1B) (Supplementary Figure S2). The sulfur content of RuBisCO from cyanobacteria containing Type IAc RuBisCO (Prochlorococcus and marine Synechococcus) was significantly (p < 0.05) lower than that of the Stramenopila (Type ID), Haptista (Type ID), and Viridiplantae (Type IB) (Supplementary Figure S2). There was no significant difference between the sulfur content of Type IAc cyanobacteria and the Rhodophyta (Type ID) (Figure 2B, Supplementary Figure S2).

Energy requirements to synthesize RuBisCO were aligned with taxonomic group and RuBisCO type (Figure 2C). It required significantly less energy to synthesize cyanobacterial RuBisCO than eukaryote RuBisCO. There was no significant difference in the amount of energy required to synthesize the two types of cyanobacterial RuBisCO (Types 1Ac and Type 1Bc) (Supplementary Figure S2). Groups containing Type 1D RuBisCO clustered together (Figure 2C), indicating that there was no significant difference between them in the amount of energy required to synthesize RuBisCO (Supplementary Figure S2). Type 1B RuBisCO, found in plants and green algae, required the most energy for RuBisCO synthesis (Supplementary Figure S2).

Data from the large and small subunits of RuBisCO were analyzed separately, which increased the number of available sequences for analysis (Table 1). Analysis of the large subunits added two supergroups that were absent from the analysis of the holoenzyme; the Alveolata and the Discoba. All the Alveolata sequences were from dinoflagellates (Dinoflagellata) and the Discoba were Euglenoidea (Table 3). Compared with all other taxa, Dinoflagellates had the longest large subunit sequences (561 ± 108; mean ± SD) (Table 3). However, there was a high degree of uncertainty in this estimate as the shortest sequence was 471 amino acids and the longest was 740 in length. The number of Alveolata sequences analyzed was small (n = 5) and there was no mode sequence length. Therefore, despite the variation in sequence length, none of the sequences could be excluded from the data based on the criteria described in the methods. There was a significant difference in the amino acid length of the large subunits between different taxa (H = 191.2, 6 degrees of freedom, p < 0.001) (Supplementary Figure S3), with the shortest sequences (471 ± 0 amino acids) occurring in the cyanobacteria with Type 1Ac RuBisCO and the longest sequences (with the exception of the dinoflagellates) occurring in the Stramenopila (490 ± 2 amino acids). Amino acid sequences were longer in the ‘red type’ lineages containing Type ID RuBisCO than the ‘green type’ lineages containing Type IA RuBisCO. There were notable differences in the numbers of specific amino acids between the different groups (Figure 3). For example, the median number of threonine in cyanobacteria containing Type IAc RuBisCO was 27, compared to 32 in cyanobacteria containing Type IBc RuBisCO. There was also variation in the number of sulfur-containing amino acids (cysteine and methionine) across the different groups. Red type Archaeplastida, containing Type 1D RuBisCO, had the minimum median number of cysteine (4) compared with the Discoba (Type IB RuBisCO), which had the maximum median number of cysteine (33). Median methionine varied from 9 in cyanobacteria containing Type IAc RuBisCO, to 18 in the red type Archaeplastida.

Pooling data from all the eukaryote groups, the mean length of the large subunit protein was 477 ± 5 (mean ± SD) (n = 403; excluding dinoflagellates). This was a small, but significant, difference from the length of the large subunit in cyanobacteria (474 ± 2; n = 166) (U = 29295, p = 0.05). The mean elemental resources required to build the large subunit in eukaryotes (excluding dinoflagellates) was 2356 ± 27 carbon, 651 ± 5 nitrogen, and 20 ± 3 sulfur atoms; which is similar to the 2361 ± 16 carbon, 652 ± 4 nitrogen, and 20 ± 3 sulfur atoms required to build the large subunit in cyanobacteria. Despite the similarity in carbon and nitrogen content of eukaryotes and cyanobacteria, there were significant differences in the median carbon (U = 23808, p < 0.001) and nitrogen content (U = 26659, p < 0.001) of these groups.

The comparative number of carbon, nitrogen, sulfur, and energy required to synthesize the large subunit varied across different taxa of photosynthetic organism (Figure 4). These data are presented at a higher taxonomic resolution than the previous data, with each data point representing a single class or division of organisms (see Table 1 for a list of classes analyzed). The classes and divisions grouped according to RuBisCO type; groups containing Type ID RuBisCO (e.g. diatoms (Bacillariophyta), golden algae (Haptista) and red algae (Bangiophyceae) required more carbon per large subunit than organisms containing both Type IB RuBisCO (e.g. green algae, land plants and cyanobacteria) and Type IA RuBisCO (Cyanobacteria) (Figure 4A, Supplementary Figure S3). Land plants (Magnoliopsida and Marchantiophyta) contained less sulfur per large subunit compared with marine groups of Eukaryotes (Bacillariophyta, Phaeophyceae and Haptista). Just as in the complete L₈S₈ RuBisCO molecule, the energy requirement to build the large subunit was proportional to its carbon content (Figure 4C, Supplementary Figure S3). There were significant differences in the elemental composition and amount of energy required to synthesize the large subunit of RuBisCO between different taxa (Supplementary Figure S3).

There was a significant difference in amino acid length of the small subunit between different taxa (H = 138.9, p < 0.001) (Figure 5, Supplementary Figure S4). Land plants and green algae (Viridiplantae) had the longest sequences, with a mean length of 177 amino acids. The mean length of the small subunit in other common groups of eukaryotes (Stramenopila, Cryptista and Hatpista) was 139 amino acids (Table 4). The mean (± SD) length of the small subunit in cyanobacteria was 112 ± 1 (n = 14) amino acids, compared with 153 ± 27 (n = 165) in eukaryotes. The Stylonematophyceae (a class of red algae) were excluded from subsequent analysis due to their extremely long sequences for the small subunit (mean = 313 amino acids) and low representation in the data (n =2). Based on the length of the small subunit in other classes of red algae, it is probable that these long sequences are a result of a sequencing error. As observed with the large subunit, there was variation in the number of specific amino acids between different groups (Figure 5). For example, the median number of alanines in Cyanobacteria containing Type IBc RuBisCO was 4 (3.6% of amino acids), compared with 18 (9.0% of amino acids) in the Viridiplantae (Type IB).

Significant variation in the length of the small subunit protein resulted in significant variation in the amount of elemental resources (C, N and S) and energy required to synthesize it across different taxonomic groups (Figure 6). There are 3 distinct clusters of taxonomic groups in Figure 6, corresponding to the Cyanobacteria (Types IAc and IBc RubisCO), eukaryotes containing Type 1D RuBisCO (Stramenopila, Haptista, Cryptista and red-type Archaeplastida (Rhodophyta)), and eukaryotes containing Type IB RuBisCO (green-type Archaeplastida (Viridiplantae)). As expected by their relatively long amino acids sequences, significantly more elemental resources in terms of carbon, nitrogen (Figure 6A), sulfur (Figure 6B), and energy (Figure 6C) were needed to build a small subunit RuBisCO in the Viridiplantae compared with the other taxa (Supplementary Figure S4). The small subunit protein of the Viridiplantae contained approximately twice as much sulfur (11 ± 3 atoms; mean ± standard deviation) as all other groups, except for the Stramenopila (8 ± 1 atoms) (Table 4, Supplementary Figure S4). The mean (± SD) elemental resources needed to build the small subunit in eukaryotes were 782 ± 127 carbon, 208 ± 35 nitrogen, and 8 ± 3 sulfur atoms. This is significantly more resources than those needed to build the small subunit in cyanobacteria; 596 ± 3 carbon (U = 28.000, p < 0.001), 151 ± 5 nitrogen (U = 28.000, p < 0.001), and 6 ± 1 sulfur atoms (U = 416.000, p < 0.001). The variation in elemental resources required to build the small subunit in eukaryotes was much greater than observed in the cyanobacteria. This was due to the significant variation in size of the small unit between different groups of eukaryotes (Table 4, Figure 5). However, the relatively small size of the sample from cyanobacteria (n = 14) compared with the eukaryotes (n = 165) may have played a role.

The mean length of the small subunit across all groups (excluding the stylonematophyceae) was 148 ± 22 amino acids (± SD; n = 177), with a range of 107 to 205 amino acids. The mean length of the large subunit across all groups (excluding dinoflagellates) was 476 ± 4 (± SD; n = 570), with a range of lengths from 465 to 496 amino acids. The lower variation in length of the large subunit compared with the small subunit suggests that its structure is more conserved. Considering just the small subunit of RuBisCO, the range of protein length was 113–113 amino acids in Type IAc, 111–113 in Type IBc, 138–141 in Type ID, and 107–205 in Type IB organisms. Variation in the Type IB was much greater than in the other groups as this group included the Viridiplantae (green algae, terrestrial plants) and Glaucophyta. Discoba (euglenoids) also contain Type IB RuBisCO, but they were not represented in the small subunit data.

The costs of creating L₈S₈ RuBisCO for the different types of RuBisCO are presented in Table 5. This is different from the data presented in Table 2, as all the groups containing Type ID RuBisCO (Bangiophyceae, Bacillariophya, Haptista, and Cryptophyceae) are grouped together. Comparing the two types of eukaryote RuBisCO, the ‘green type’ (Type IB) contains significantly more carbon and nitrogen compared with the ‘red type’ (Type ID) (Table 5, Supplementary Figure S5). There were significant differences in the C:N ratio (F_3,37 = 11.240, p < 0.001) of the different types of RuBisCO, with Type ID having a significantly (p < 0.05) higher C:N ratio than Type IB and Type IBc (Table 5, Supplementary Figure S5). The C:N ratio of the two types of cyanobacterial RuBisCO were also significantly (p < 0.05) different, with Type IAc (Prochlorococcus and marine Synechococcus) having a higher C:N ratio than Type IBc RuBisCO. There was no significant difference in the C:S or N:S ratios of the different types of RuBisCO (Table 5, Supplementary Figure S5). The number of nitrogens per holoenzyme was a much more useful metric for comparing RuBisCO between groups than the C:N ratio. This is simply because there is significant variation in the number of amino acids required to build RuBisCO between the different Types (Table 5). For example, while both Types IAc and ID have a mean C:N ratio of 3.71, it takes a mean 6,363 nitrogen atoms to build the holoenzyme of Type IAc RuBisCO compared with 6,765 nitrogen atoms for Type ID.

The mean energy: C ratio for the four RuBisCO types was constant (5.84 to 5.86 high energy phosphate bonds per carbon), reflecting the fact that it is the carbon content that determines how much energy is required to build the proteins and it is not affected by the nitrogen and sulfur content of the RuBisCO. For example, the cost of synthesizing glycine (2 carbons) is 11.7 phosphate bonds, compared with 74.3 phosphate bonds for tryptophan (11 carbons) (Supplementary Table S1). There was a significant correlation (r = 0.997, p < 0.001, n = 41) between the carbon content of the protein and the estimated energy required to synthesize it.

The largest difference between the elemental composition of the RuBisCo protein corresponds to the difference between cyanobacteria and eukaryotes (Table 2). It requires ~ 390 more amino acids to build eukaryote RuBisCO compared to cyanobacterial RuBisCO. The resources required to build an L₈S₈ RuBisCO in eukaryotes were approximately 1,900, 400 and 37 more carbon, nitrogen and sulfur atoms, respectively, than in cyanobacteria. This finding fits with the observation that proteins are significantly longer in Eukarya compared with Bacteria (Zhang, 2000; Brocchieri and Karlin, 2005; Tiessen et al., 2012). Tiessen et al. (2012) found that the average protein length in Eukarya was 472 amino acids, compared with 320 in Bacteria and 283 in Archaea. By these criteria, the small RuBisCO subunit is smaller than an average protein in both eukaryotes and cyanobacteria (Table 4). The large RuBisCO subunit is larger than the average protein in cyanobacteria and slightly larger than average in eukaryotes (Table 3).

A dimer of the large subunit (L₂) is the minimum active unit able to fix carbon in Form I RuBisCO (Tabita et al., 2008a). While the large subunit is essential for the function of RuBisCO, the small subunit is hypothesized to play a structural role in stabilizing the holoenzyme and moderating its catalytic efficiency (Mao et al., 2023; Amritkar et al., 2025). This may explain why there was much more variation in protein length in the small subunit protein than the large subunit; the large subunit must be relatively conserved to maintain the active site and its catalytic function when organized in a dimer and activated.

Type 1B RuBisCO is found in both multicellular and single celled organisms, which inhabit marine, freshwater, and terrestrial habitats. Variation in the length of the small subunit did not show patterns associated with specific habitats. Even within one family associated with one environment, there was considerable variation in small subunit length. The Mamiellophyceae had small subunit protein lengths of 162 to 204 amino acids and all species analyzed were marine. The small subunit length in land plants (Magnoliopsida) varied between 177–183 amino acids, similar to those of marine green macroalgae such as Ulva. Variation in the length of the small subunit in land plants may be due to multiple isoforms of rbcS within a single plant species, which are expressed under different environmental conditions (Dedonder et al., 1993). The freshwater glaucophytes (Cyanophora spp.) were outliers with the shortest small subunit proteins (107 amino acids). Glaucophyta are a primitive group that diverged from the Viridiplantae early in the evolutionary history of the Archaeplastida, around 1500 to 1600 Ma (De Clerck et al., 2012; Leebens-Mack et al., 2019). While variation in overall length may be challenging to relate to enzyme function, specific regions of the protein are known to be important, such as the βA-βB loop. Green algae have long (20–31 amino acid residues) βA-βB loops compared with plants (∼ 22 amino acid residues) and non-green algae and prokaryotes (∼ 10 amino acid residues) (Mao et al., 2023). When the βA-βB loop in the green alga Chlamydomonas reinhardtii was substituted for the shorter βA-βB loop of spinach, the resulting RuBisCO from the mutant C. reinhardtii had carboxylation and oxygenation kinetic constants indistinguishable from spinach (Spreitzer et al., 2005).

Pooling the data for all the complete holoenzymes analyzed, the C:N ratio was 3.69 ± 0.04 (mean ± SD; n = 41). The general stoichiometric equation for phytoplankton proteins is C₁₀₆H₁₆₈O₃₄N₂₈S (Sarmiento and Gruber, 2006), or a C:N ratio of 3.79. Inomura et al. (2020) report a C:N value for phytoplankton protein of 3.82, based on the work of Brown (1991). These values suggest that RuBisCO is enriched in nitrogen when compared with average phytoplankton protein (Table 5). Therefore, RuBisCO must contain a relatively high proportion of amino acids that contain 2 or more N atoms (arginine, asparagine, glutamine, histidine, lysine, and tryptophan; Supplementary Table S1).

Baudouin-Cornu et al. (2001) found significant correlations between atomic composition and metabolic function – sulfur and carbon assimilatory enzymes were depleted in S and C, respectively, in both Escherichia coli (a bacterium) and Saccharomyces cerevisiae (a yeast). The enrichment of nitrogen relative to C in RuBisCO suggests a similar pattern as the primary function of RuBisCO is the acquisition of reduced carbon during photosynthesis. This observation fits with a stoichiogenomic paradigm in which natural selection has biased monomer usage (i.e. amino acids in proteins) to reduce elemental costs associated with limiting elements (Elser et al., 2011). While there were differences between the C:N ratio between different supergroups of phytoplankton, these differences are not as important as the variation in protein length in determining nitrogen and carbon requirements.

The sequence lengths of the large subunit in the dinoflagellates were exceptionally long (561 ± 108 amino acids; mean ± SD) compared with the next longest group, the Stramenopila (490 ± 2). This may have been due to fundamental differences between the large subunit protein of dinoflagellates compared with all the other photosynthetic organisms. More likely, the wide range of protein lengths suggests that there were errors in the sequencing from dinoflagellates and therefore the data were unreliable. This was compounded by the very small (n = 5) sample size and consequently no modal value, so outliers could not be eliminated.

The organization of genetic material in dinoflagellates is very different from that in other eukaryote phytoplankton and they are challenging to sequence (Bachvaroff and Place, 2008; Beauchemin et al., 2012; Lin, 2024). In addition, there are fundamental differences between the structure and organization of RuBisCO in dinoflagellates compared with the other major groups of marine phytoplankton. Dinoflagellates contain Form II RuBisCO encoded in the nucleus, whereas all other eukaryote groups in this work contain Form I RuBisCO in which the large subunit genes reside in the plastid and the small subunit genes reside inside the nucleus (Morse et al., 1995; Rowan et al., 1996; Tabita et al., 2008b; Gruber and Feiz, 2018). Dinoflagellate RuBisCO is fundamentally different from the hexadecameric (L₈S₈) structure of Form I RuBisCO. Dinoflagellates RuBisCO is a homodimer of the large subunit (L₂), or in the form (L₂)_n where n is an integer of 5 or less (Tabita et al., 2008b; Gruber and Feiz, 2018; Rydzy et al., 2021). This explains why there were no small subunit sequences for dinoflagellates in these data (Table 1).

This study has considered resources in terms of carbon, nitrogen, sulfur, and energy in the form of phosphate bonds. As the stoichiometric equation for phytoplankton protein shows (C₁₀₆H₁₆₈O₃₄N₂₈S; Sarmiento and Gruber, 2006), Oxygen and hydrogen are very abundant elements in proteins. Using Type 1D RuBisCO as an example, it has a mean oxygen content of 7,368 atoms and 38,854 hydrogen atoms, compared with 6,765 nitrogen and 25,090 carbon atoms. Despite their abundance, oxygen and hydrogen data were not presented in this paper as they are not limiting elements in the ocean, and the energy required to incorporate them into amino acids was included in the energy cost of synthesizing the amino acids.

Energy costs were presented in terms of phosphate bonds, but this does not represent a consumption of phosphorus due to the rapid recycling of phosphate through adenosine diphosphate (ADP) (C₁₀H₁₅N₅O₁₀P₂) and back to ATP (C₁₀H₁₆N₅O₁₃P₃) within the cell. Nevertheless, a large pool of P is needed within cells to maintain metabolism. Based on cellular ATP measurements by Hamilton and Holm-Hansen (1967), the amount of P associated with ATP in marine bacteria ranges from 2.96 x 10^–3 to 38.4 x 10^–3 fmol P cell^-1. For aquatic eukaryotes (both marine and freshwater), with cell diameters from 10 to 50 μm, the amount of P in ATP ranges from 0.946 to 118 fmol P cell^-1 (Peperzak et al., 2024).

The number of carbon atoms dictated the energetic cost of synthesizing the amino acids as the number of nitrogen or sulfur atoms in the amino acid did not significantly add to the cost (Supplementary Table S1, Supplementary Figure S1). The cost associated with polymerization of amino acids was proportional to the number of amino acids, and therefore carbon atoms. As a result, the energy cost of synthesizing RuBisCO was highly correlated with its carbon content. Calculating the energy costs of synthesizing RuBisCO provided no more insight than counting the number of carbon atoms in the protein. A more sophisticated approach to evaluate the energy cost of synthesizing RuBisCo would include indirect costs. Examples of indirect energy costs include those associated with the acquisition of biologically available nitrogen and phosphorus from the environment, transcription costs, and the costs associated with synthesizing the chaperon proteins required to synthesize the holoenzyme (Gruber and Feiz, 2018).

Time can also be considered a resource as it takes more time to obtain the resources and synthesize longer proteins. Protein elongation in a range of organisms (both bacteria and eukaryotes) is 0.59 to 21 amino acids per ribosome per second (Karpinets et al., 2006). Using the mean length of Type IB RuBisCO as an example (5229 amino acids in the L₈S₈ holoenzyme), this would be equivalent to a time range of 249 to 8,863 seconds (4 to 148 minutes). This compares to 222 to 7,905 (4 to 132 minutes) for Type IBc RuBisCO. As RuBisCO has a long half-life (5 to 9 days in Zea mays; Simpson et al., 1981), these time differences may not be significant, or act as a constraint, in maximizing photosynthesis. However, it does mean that the ribosomes are not available to synthesize other proteins. Of course, protein elongation rates are one of many time costs in synthesizing a protein, some of which are directly proportional to the lengths of the protein (e.g. mRNA synthesis) and others that are not (initiation and termination).

The translation of proteins encoded in genes does not necessarily result in the final mature protein due to modifications during translation (co-translational modifications) and at any point during the lifetime of the protein after translation (post-translational modifications (PTMs)) Kessel and Ben-Tal, 2018; Morales-Polanco et al., 2022; Zhong et al., 2023). Therefore, the elemental compositions of the mature protein will differ slightly from that of the encoded sequence. Even the composition of the mature protein is not fixed and may change as the organism responds to the environment (Kessel and Ben-Tal, 2018; Amaral et al., 2024). The number of known PTMs in protein are > 650 and these are generally poorly characterized in specific proteins (Zhong et al., 2023). Therefore, there are not a fixed set of PTMs that can be accounted for in these calculations. It is known that RuBisCO is controlled by several PTMs, including phosphorylation, tyrosine-nitration, acetylation, lysine-methylation, nitrosylation (NO) and glutathionylation (Grabsztunowicz et al., 2017). The large subunit of diatoms is known to undergo a number of PTMs, such as 4-hydroxyproline, ß-hydroxyleucine, hydroxylated and nitrosylated cysteine, mono- and dihydroxylated lysine, and trimethylated lysine (Valegård et al., 2018). A consistent PTM in RuBisCO is the carbamylation of an active-site lysine residue, which is essential for RuBisCO activity as it is needed for RuBisCO to bind to its substrate, ribulose-1,5-bisphosphate (RuBP) (Spreitzer and Salvucci, 2002; Valegård et al., 2018). The calculations presented in this paper assume that these modifications of amino acid residues did not have a significant impact on the elemental composition of RuBisCO and that what is encoded in the genes is at least an accurate representation of the nascent or immature protein.

Bar-On and Milo (2019) estimated that the ocean contains a standing stock of phytoplankton that contain approximately 0.03 Gt of RuBisCO, assuming that RuBisCO is 3% of the protein content of the cells. Both the RuBisCO and protein content of phytoplankton are not fixed, depending on genotype and how organisms respond to the environment to affect phenotype (Losh et al., 2013; Jónasdóttir, 2019). Bar-On and Milo (2019) combined 10 reports on the RuBisCO content of microalgae. The geometric mean of these values, plus an assessment of the relative biomass of different phytoplankton groups in the ocean (from Bar-On et al., 2018), was used to estimate 0.03 Gt (Bar-On and Milo, 2019). It should be acknowledged that the 0.03 Gt estimate is based on limited measurements of the cellular content of RuBisCO in phytoplankton. The molecular mass and elemental content data presented in Table 5, in addition to the estimated 0.03 Gt RuBisCO in the ocean, was used to estimate the amount of biogeochemically significant elements (C, N and S) associated with RuBisCO. There is approximately 0.016 Pg C, 5.1 Tg N, and 0.4 Tg S in the form of phytoplankton RuBisCO in the ocean. There was no significant difference between the different Types of RuBisCO in this simple calculation as there were more moles (5.71 x 10⁷ moles) of Type IAc RuBisCO in 0.03 Gt compared with Type 1B (5.16 x 10⁷ moles), thus the difference between the elemental composition of individual RuBisCO holoenzymes was counteracted by the relative amount of each enzyme in the calculation. The same method can be used to estimate the annual productivity of RuBisCO. Estimates for the annual productivity of the ocean are 45 to 55 Pg C yr⁻¹ (Longhurst et al., 1995; Field et al., 1998; Falkowski et al., 2000; Carr et al., 2006; Westberry et al., 2008), which using the conversion of Bar-On and Milo (2019), results in 1.35 to 1.65 Gt RuBisCO yr^-1. This is equivalent to 0.725 - 0.890 Pg C, 228–283 Tg N, and 16.5 - 22.5 Tg S fluxing through the RuBisCO protein each year.

Bioavailable nitrogen is a limiting nutrient in a significant proportion of the surface ocean (Moore et al., 2013; Browning and Moore, 2023), with nitrogen supply constrained by the amount of biological nitrogen fixation (a process which is often limited by the availability of iron), and the supply of bioavailable nitrogen from the land. Estimates of the amount of nitrogen fixed by cyanobacteria and other microorganisms in the ocean is 223 ± 30 Tg N yr⁻¹ (Shao et al., 2023). Therefore, amount of bioavailable nitrogen fluxing through the pool of RuBisCO in the ocean is equivalent to, or even higher, than the total annual input of bioavailable nitrogen from nitrogen fixation. As the standing stock of nitrogen in RuBisCO is 5.1 Tg N, the turnover of the RuBisCO pool must be rapid, occurring every 6.6 to 8.2 days. This is similar to estimates of 2 to 6 days for the turnover of phytoplankton biomass (Behrenfeld and Falkowski, 1997).

Prochlorococcus is considered to be the most numerically abundant photosynthetic organism on Earth and has a significant impact on biogeochemical cycling in the ocean (Biller et al., 2015; Flombaum et al., 2013; Partensky et al., 1999; Visintini et al., 2021). The total amount of RuBisCO in Prochlorococcus was estimated using data in Table 5 and published information. The estimated abundance of Prochlorococcus is 2.8 x 10²⁷ to 3.0 x 10²⁷ cells in the ocean at any one time (Flombaum et al., 2013). The RuBisCo content of Prochlorococcus is 7.55 x 10^–22 to 8.954 x 10^–22 moles of holoenzyme (L₈S₈) per cell (Hopkinson et al., 2014). Therefore, the estimated amount of RuBisCO in Prochlorococcus is 2.11- 2.69 x 10⁶ mol, which amounts to 4 to 5% of the total RuBisCO pool in the ocean. This RuBisCO contains 6.0 – 7.6 x 10^–4 Pg C, 0.18 – 0.24 Tg N, and 0.014 – 0.017 Tg S. Prochlorococcus fixes 4 Pg C yr^-1 (Flombaum et al., 2013), or 7.3 – 8.9% of net ocean primary productivity. The relatively low proportion of RuBisCO in Prochlorococcus compared with its net productivity suggests that it is highly efficient at photosynthesis. This conclusion, based on global scale biogeochemical observations, fits with what we know about the physiology of Prochlorococcus. It has CO₂ fixation rates 1.5 to 2 times higher than those of Synechococcus and photosynthetic picoeukaryotes when rates are normalized to photosynthetic pigment concentrations (Hartmann et al., 2014). This is due to adaptations such as a highly efficient carbon concentrating mechanism to elevate the concentration of CO₂ around RuBisCO (Hopkinson et al., 2014).

The timing of the emergence and/or dominance of different groups of phytoplankton is coupled to the geochemical history of the Earth (Falkowski et al., 2004; Katz et al., 2007). Earth’s biogeochemistry has changed significantly since the emergence of oxygenic photosynthesis, with many of the changes driven by the activity of the photosynthesizers themselves. The oxygenation of the ocean and atmosphere is the obvious example (Lyons et al., 2024; Stockey et al., 2024). While RuBisCO existed before the GOE, the origin of Form I RuBisCO coincided with the GOE, so it is ~ 2.5 billion years old (Amritkar et al., 2025). RuBisCO is one of the slowest evolving enzymes on Earth, with a nucleotide substitution in the rbcL gene every 0.9 My and one amino acid mutation every 7.2 My (Bouvier et al., 2024). Since the GOE there has been time for approximately 139 amino acid substitutions, equivalent to a substitution in approximately 30% of the amino acids in the large subunit. This is sufficient to account for the observed differences between the large subunit in different major groups of phytoplankton (Figure 3).

There have been many global scale transitions and events that have influenced the evolution of phytoplankton over the last 2.5 billion years, including the ‘big five’ mass extinction events, and snowball Earth periods (Knoll, 2004; Lenton and Watson, 2011). However, there have also been several secular trends that have affected resource availability. These will have affected the evolution and dominance of different phytoplankton groups and their proteins. The Sun’s luminosity has increased by 30% over the course of Earth’s 4.6 billion year history (Lenton, 2016). The Sun’s luminosity increased by 15 – 20% over the Proterozoic Eon (2,500–541 Ma), and by a further ~ 4% over the Phanerozoic Eon (last 541 Ma) (Tajika, 2003; Basinger et al., 2024). This has resulted in significant increase in the flux of solar radiation available to drive photosynthesis at the Earth’s surface, on the order of 20 – 25% over the history of Form I RuBisCO. Phosphorus is regarded as the ultimate limiting nutrient on geological timescales (Tyrrell, 1999). There has been considerable debate over the concentrations and bioavailability of phosphorus in the Precambrian ocean, which are beyond the scope of this paper (e.g. see Robbins et al., 2016). However, there seems to have been a shift in the phosphorus cycle in the late Proterozoic (Planavsky et al., 2010, 2023; Reinhard et al., 2017), resulting in an increase in the bioavailable phosphate pool. Planavsky et al. (2023) times this shift at ~ 750 Ma in the Tonian period (1000–720 Ma). The pool of bioavailable nitrogen in the ocean has increased significantly over the last 2.5 billion years. Biological nitrogen fixation probably existed before oxygenic photosynthesis (Stüeken et al., 2015, 2024), but bioavailable nitrogen (ammonium) did not accumulate in the anoxic and Fe²⁺ rich ocean as it was coupled to processes such as iron-ammonium redox reactions (Feammox) that returned nitrogen to the atmospheric sink as N₂ (Stüeken et al., 2024). The oxygenation of the ocean resulted in a more complex nitrogen cycle, in which nitrification played a major role, resulting in a pool of bioavailable nitrate (Stüeken et al., 2024). Nitrate availability remained low and doubled around 800 million years ago, which coincides with the shift to eukaryote dominated ecosystems (Lenton et al., 2014; Kang et al., 2023). The availability of sulphur in the ocean has been strongly affected by the oxidative state of the Earth’s surface. Over the history of life on Earth, there has been an increase in the sulfate concentrations in the ocean from µM in the Archean to concentrations in the µM to low mM after the GOE (Paleoproterozoic and Mesoproterozoic), to concentrations of 28 mM today (Fakhraee et al., 2019, 2025; Lyons et al., 2024).

These changes collectively show an increase in resources (light energy, bioavailable N and P, and S) essential for building phytoplankton. The increase in bioavailable N and P in the Tonian Period (1000 to 720 Ma) of the Neoproterozoic Era was a driver in the rise to dominance of eukaryote phytoplankton and increased productivity during the Neoproterozoic and subsequent Paleozoic Eras (1000 to 252 Ma) (Brocks et al., 2023; Kang et al., 2023; Stockey et al., 2024). The shift to eukaryote phytoplankton led to an increase in the proportion of larger proteins in the phyoplankton metaproteome that require more energy, C, N, S, and P to synthesize. Eukaryote RuBisCO contains ~ 8% more carbon, nitrogen, and energy per molecule compared with cyanobacterial. This represents a significant increase as protein contains about half of the carbon in a phytoplankton cell, and 2-23% of the protein mass is RuBisCO (Bar-On and Milo, 2019; Ellis, 1979; Losh et al., 2013). While it takes ~ 13% more sulphur to synthesize eukaryote RuBisCO compared with cyanobacterial, it is unlikely that sulphur was ever a limiting nutrient for marine phytoplankton as sulphate concentrations have probably been > 0.1 mM since the GOE (Fakhraee et al., 2025). The increase in resource availability created conditions where the competitive cost of synthesizing larger proteins became less significant for eukaryote phytoplankton competing with cyanobacteria.

The red lineages (diatoms, coccolithophores, and dinoflagellates) of eukaryote phytoplankton that dominate the modern ocean today came to prominence during the Mesozoic Era (251 to 65 Ma) (Falkowski et al., 2004). Green eukaryote phytoplankton dominated before the Mesozoic. Rickaby and Hubbard (2019) proposed that the transition from green to red eukaryote phytoplankton was driven by the different properties of Type IB and Type ID RuBisCO, and the increasing oxygenation of the ocean during the Paleozoic Era (541–252 Ma). Type IB RuBisCO, characteristic of chlorophyte phytoplankton, has relatively poor selectivity between O₂ and CO₂, therefore it is suited to lower oxygen oceans with lower O₂ to CO₂ ratios. Type 1D RuBisCO is better adapted to higher O₂ to CO₂ ratios and therefore more oxygenated oceans. Red lineages with Type ID RuBisCO were able to outcompete green eukaryotes as they did not need to invest as many resources in carbon concentrating mechanisms as the ratio of O₂ to CO₂ increased (Rickaby and Hubbard, 2019).