The endosymbiont genomes to be included in the analyses and their accession IDs were mostly retrieved from SymGenDB (Reyes-Prieto et al., 2020). We added the following genomes that were not available in SymGenDB (2020) due to their posterior discovery or annotation, or because they are not listed in the KEGG database: “Candidatus Serratia symbiotica” SeCistrobi, “Candidatus Tremblaya phenacola” PPER, “Candidatus Tremblaya princeps” TPPLON, Cardinium cSfur, Cardinium hertigii cBtQ1, Neisseria meningitidis MC58, Serratia symbiotica SCt-VLC, “Candidatus Tremblaya princeps” TPMHIR1, “Candidatus Tremblaya princeps” TPPMAR1, “Candidatus Tremblaya princeps” TPFVIR, “Candidatus Tremblaya princeps” TPTPER1, “Candidatus Sulcia muelleri” TETUND, “Candidatus Hodgkinia cicadicola” TETUND2, Buchnera aphidicola BCc and “Candidatus Sodalis sp.” SoCistrobi. Their genomic data were retrieved automatically on August 2020 using the efetch command from GeneBank,¹ except for “Candidatus Tremblaya phenacola” PPER, Serratia symbiotica SCt-VLC and Cardinium hertigii cBtQ1, which were manually downloaded because only shotgun assemblies are available. Additionally, in order to generate a complete universe of translation-involved genes, we included in our analysis 10 bacterial genomes with no reduction, two of them, E. coli and B. subtilis, are common bacterial models for Gram-negative and Gram-positive bacteria, respectively. The other eight were selected because they are taxonomically diverse and can be grown in the laboratory in axenic culture. The Prokka software tool (Seemann, 2014) was used to re-annotate all genomes for homogeneous results. Finally, we also included in our comparisons the genome of M. mycoides JCVI-syn3.0. This genome annotation was retrieved from Hutchison et al. (2016). We only took into consideration genes classified as ribosome biogenesis, RNA metabolism, protein folding, translation, RNA, rRNA modification, tRNA modification, and regulation. The sequences of ORFs classified as “unclear category” were used to perform a BLASTP against the non-redundant protein sequences database at the NCBI web (The Blast Sequence Analysis Tool, 2022) to look for putative functions of the hypothetical conserved proteins they might encode. All 110 bacterial genomes under study have been compiled in a dataset called “cosym” with 92 entries as a result of considering coprimary insect endosymbionts (symbiotic consortia) as single entities (Supplementary Material 1), and a list of the genera included in this study is listed in Table 1.

Genes were classified into functional categories based on the previous work by Grosjean et al. (2014). We defined the set of translational protein-coding genes (translational gene set) of the two model species considered through several steps. First, we retrieved the genes from E. coli and B. subtilis that have been included in the work by Grosjean et al. (2014) as encoding essential components of the translation machinery. Then, we searched for selected GeneOntology (GO) terms on UniProt (UniProt Consortium, 2015) and EcoCyc (Keseler et al., 2013), in order to update the list (Supplementary Material 2). The GO terms included were: 0000154 (rRNA modification), 0001510 (RNA methylation), 0001680 (Addition of CCA 3’-end of tRNA), 0005840 (Ribosome), 0006364 (rRNA processing), 0006396 (RNA processing), 0006400 (tRNA modification), 0006412 (Translation), 0006417 (Translational regulation), 0006457 (Protein folding), 0008033 (tRNA processing), 0009451 (RNA modification), 0042255 (Ribosomal assembly), and 0042254 (Ribosomal biogenesis). In this step, a manual curation of gene names was mandatory to eliminate duplicated candidates (i.e., the same ortholog with different annotated names), to use the UniProt accepted nomenclature for genes with a double translational function, and to remove genes not strictly related to translation.

Orthologous genes (paralogs included) in all genomes under study were identified using the Roary software (Page et al., 2015) with default parameters. The absence (0 count) or presence (1 count) of each one of these orthologs in each genome was counted, creating several matrices for further analyses, one per each functional translation category (Supplementary Material 3).

Hierarchical cluster analysis (HCA), an unsupervised machine learning approach for grouping datasets into clusters, was used for the classification of the dataset genes. HCA was performed in R with gplots:heatmap.2 and stats:hclust (complete method) functions. The input data were the presence/absence matrices obtained in the gene orthology analysis. To extract the gene names of the dendrogram clustering, the cutree function of the stats R-package was used. The treatment of the data, the construction of figures and the statistical analyses performed for this work were carried out with ad doc scripting in RStudio 4.0.3 using the stats, ggplot2 (Wickham, 2016) and ggthemes R packages. The full script is available through https://github.com/majogarzon/MinTransMach.git.

Defining the basic living functions to generate a simplified bacterium that might be modulated under laboratory conditions with desirable and predictable outcomes for biotechnological purposes is a great challenge. The naturally reduced genomes of insect endosymbionts have historically been studied to approach the minimal genome concept, providing valuable information about those functional modules that are necessary and sufficient for life. Previous comparative studies (Gil et al., 2004) concluded that the minimal genome is substantially enriched in genes involved in genetic-information processing, mainly coding for the elements of the translational apparatus, the most complex machinery in a living cell. Yet, it has been possible to define simplified but still functional translation machineries after a reductive evolution in Mollicutes and insect endosymbionts (Grosjean et al., 2014; Gil and Peretó, 2015).

In this work, to further explore and validate the minimal translation machinery, our search began with the selection of organisms with naturally reduced genomes to compare them with known free-living bacterial models which must possess efficient, complete, and more complex translational apparatus. As a starting point, taking advantage of the availability of extensive genomic information and the bioinformatic tools developed in recent years, we retrieved all the genomes annotated as insect endosymbionts from SymGenDB, a database that lodges genomic information of organisms involved in symbiotic relationships (Reyes-Prieto et al., 2020), and updated the information by manually including some additional endosymbiont genomes, as described in Materials and Methods. All these genomes were classified depending on their symbiotic relationship as primary or obligate symbionts (OS) when they are necessary for the survival and reproduction of the host, and as secondary symbionts (SS) when they maintain a facultative symbiotic relationship in terms of survival. Furthermore, we included several non-symbiotic organisms, designated as free-living (FL), to have a complete representation of the universe of translational genes in our data set. All the 110 organisms used in our analyses are listed along with taxonomic and genomic information in Supplementary Material 1, and a summary of their genera is listed in Table 1.

Many insects live in obligate association with more than one endosymbiotic bacterium. It has been observed that the presence of two (or more) co-primary endosymbionts allows a greater reduction of the bacterial genomes, far below the definition of a minimal genome (Sloan and Moran, 2012). Probably this means that they can complement each other by exchanging some gene products to perform essential functions (Reyes-Prieto et al., 2014), including informational ones (i.e., DNA replication, transcription and translation). For this reason, to generate our final genome dataset we considered all coprimary symbionts as a single entity (consortia sheet in Supplementary Material 1), leading to only 92 entries. Next, we searched for the orthologs in the genome’s dataset using the Roary software. We found 101,012 clusters of orthologs which compose the whole pangenome of the organisms under study.

Before determining our universe of translational genes, we had to cope with the existence of poorly or wrongly annotated genes and pseudogenes within the analyzed genomes. In fact, a critical difficulty in carrying out this work has been the automation of the process because, even though great efforts are being made to unify the nomenclature (as defined by the International Nucleotide Sequence Database Collaboration, INSDC; Brunak et al., 2002), at present no database provides the unified names for all genes. Most gene descriptions are still based on their initial identification by classical genetics in each given organism, and a recent discovery of their function is often associated to errors, such as entering a function twice without unified descriptors or simply by not including it in databases. For this work, the UniProt nomenclature recommended by the INSDC has been used; in cases where there was no classification, the accepted nomenclature for in E. coli K-12 MG1655 (taxonomy ID 511145) was chosen (Schoch et al., 2020).

Once this problem was solved, we set up a universe of translational protein-coding genes, which is composed by the genes that encode the proteins that integrate the translation apparatus, as well as those that directly or indirectly participate in the different stages of rRNA or tRNA processing. This gene set was defined based on the genomes of the two selected model bacteria, E. coli K-12 MG1655 and B. subtilis 168. We detected a total of 309 unique genes (Supplementary Material 2). Most of them belong to the ribosomal proteins and tRNA modification categories, and only a few of them to RNA processing (Figure 2A).

Then, we compared the previously defined pangenome to the universe of translational genes, to search for the 309 genes involved in translation. As expected, the genome reduction process affects the total number of genes for this function, and the losses depend on the translational subprocess involved (Figure 2B). Most ribosomal proteins are present in all organisms regardless of their lifestyle, confirming the importance of the whole ribosome as a functional unit. In contrast, many genes of the other translational categories have been lost, especially in OS and, to a lesser extent, in SS, certainly affected by the genome reduction syndrome. The less conserved genes belong to the ribosome assembly and rRNA/tRNA modification categories, as previously described (de Crécy-Lagard, 2007; Grosjean et al., 2014).

Next, we explored the datasets for essential and persistent genes (Acevedo-Rocha et al., 2013) to get closer to the definition of a functional minimized translation machinery. Evidently, if all organisms have an ortholog of a given gene, its function must be essential and must be included in the minimal translation machinery. Conversely, a gene that only presents orthologs in few organisms with reduced genomes, would not be essential. Based on this notion, a hierarchical clustering analysis was carried out, grouping the data in each of the seven translational subprocesses considered. This machine-learning unsupervised classification algorithm managed to separate each subset of data into two clusters (Supplementary Figure 1). A total of 134 orthologs were found in the seven clusters with higher counts and represent the candidate genes to be included in a first proposal of a minimal translational machinery (Supplementary Material 4). In previous works, Buchnera aphidicola BCc the OS of the aphid Cinara cedri (Pérez-Brocal et al., 2006; abbreviated as bcc in our study) was considered to possess a small genome close to what could be considered a minimal genome, still able to support the translation process autonomously, while tiny genomes were those that had already lost this ability, so that even genes essential to the process had been lost, making them dependent on the cooperation of a cosymbiont, or even the host, to perform translation (Reyes-Prieto et al., 2014). Therefore, as a first proxy for validation of our approach to define a minimal translational machinery, we compared our results with the genes from the bcc genome as a naturally minimized reference genome. The bcc genome retains around half of the universe of translational orthologs (150 out of 309 genes), of which it shares 125 genes with our first minimal proposal. Additionally, our proposal contains nine genes (queA, rlmB, rne, rnhA, rplR, tgt, trmB, tsaC, and tusE) that are not present in bcc. As bcc is cosymbiont of Serratia symbiotica str. “Cinara cedri” (SS of Cinara cedri; abbreviated ssz), we searched for those genes in the cosymbiont’s genome. The presence of all nine genes in ssz indicates that it probably contributes essential translation genes to the symbiotic relationship. On the other side, the 25 additional genes present in bcc and absent in our minimal proposal might reflect that its genome reduction is still an ongoing process. However, as 14 out of these 25 genes are essential in E. coli (see next section) and belong to all possible translational subcategories, it cannot be ruled out that some of them are necessary to improve the efficiency of translation in the specific intracellular environment of this bacterium.

In order to test the viability of our first proposal, we compared it with the set of essential genes involved in translation of E. coli K-12 MG1655 and B. subtilis 168, according to the DEG database (Luo et al., 2021) and more recent studies on essentiality in B. subtilis (Koo et al., 2017; Pedreira et al., 2022) and E. coli (Goodall et al., 2020), and with the synthetic JCVI-syn3.0 organism (Hutchison et al., 2016). These two comparisons provide complementary information. While the DEG database reports those genes indispensable for the immediate survival of an organism, the information obtained from JCVI-syn3.0 also highlights the importance of persistent genes, needed for the cell to maintain itself for an extended term. Hutchison et al. (2016) indicate that the JCVI-syn3.0 genome has 195 genes for genetic information storage and processing. Among them, we found that 144 genes are involved in translational processes. Additionally, because 92 genes without assigned biological function were predicted when the JCVI-syn3.0 genome was published (Hutchison et al., 2016), we searched for putative functions of this last set of genes by BLASTP against the non-redundant protein sequences (nr) database from the NCBI web page. Several possibly interesting enzymes were found: a bifunctional oligoribonuclease and PAP phosphatase (nrnA; EC 3.1.3.7), a putative pre-16S rRNA nuclease RNaseH-like (yqgF; EC 3.1.-.-) and a ribosomal L7Ae/L30e/S12e/Gadd45 family protein (EC 3.1.26.5). The later must be a rpmD-like gene, as no rpmD has been annotated on the JCVI-syn3.0 genome, although it is present in all minimal sets we are working with, an indication that it must be essential. Moreover, a putative duplicated pheT gene was found. Figure 3 shows the comparison among the three translational datasets. All three datasets, share 95 genes, thus confirming that not all the genes needed are strictly essential (Acevedo-Rocha et al., 2013). Finally, of the 147-genes identified as components of the JCVI-syn3.0 translation machinery, 106 genes (72.1%) are shared with our new proposal for minimal a translation machinery and correspond mainly to ribosomal proteins and aminoacylation enzymes (Supplementary Material 4).

Lastly, based on information from the BioCyc and UniProt databases, we performed a functional analysis of the genes found in JCVI-syn3.0 that had not been included in our first proposal, to improve the efficiency of the translational machinery model. Furthermore, through this functional analysis we checked if any function from our first proposal was unnecessary or redundant. For example, in the cases in which there were specific genes of Gram-negative or Gram-positive bacteria (i.e., non-orthologous gene displacement), since this study is based mainly on data from Gram-negative bacteria, we chose to include the alternative corresponding to this group. Thus, regarding aminoacyl-tRNA synthetases (EC 6.1.1.-), our proposal has included glnS and gltX instead of gluS, gatA, gatB, and gatC. As for the tRNA modification genes, we think that only thiI and tsaE need to be added. Both are related to the modifications at position 8 and 37 of tRNA, respectively. ThiI (EC 2.8.1.4) is an enzyme that produces 4-thiouridine [s(4)U8] (Kambampati and Lauhon, 2000), and it is encoded by an essential gene in E. coli K-12 MG1655 (Rajakovich et al., 2012). Moreover, IscS-IscU (EC 2.8.1.7 and EC 3.6.4.10, respectively), two partners of ThiI involved in biological iron–sulfur cluster assembly, needed for sulfur transfer, are already included in our proposal, are essential in both E. coli and B. subtilis, and are present in JCVI-syn3.0 genome. Finally, we consider that iscA should be included as well, although it is not in any of the results shown so far (i.e., JCVI-syn3.0, DEG or our first proposal). This decision is because IscA is necessary for the proper operation of the IscS-IscU system. Although any other alternative enzyme of the HesB family could replace it (López-Madrigal et al., 2013), at least one of them must be part of the minimal translation machinery. On the other hand, TsaE is involved in the formation of a threonylcarbamoyl group (t₆A37), a universally conserved modification (Thiaville et al., 2015). It acts with TsaB, TsaC (EC 2.7.7.87) and TsaD (EC 2.3.1.234), encoded by essential genes included in DEG, and present both in our first proposal and in the JCVI-syn3.0 genome. Regarding to the subcategory translation, we added tufA and tufB genes, both responsible for the formation of the EF-Tu protein whose essential function is delivering aminoacylated tRNA into the A-site of the ribosome during protein biosynthesis (Kacar et al., 2017). In addition of being necessary for proper translation, they are considered essential in the DEG databases, and are present in the JCVI-syn3.0 genome. As for the subcategory Ribosome assembly, we think that rimM must be added because, together with rbfA, is needed for efficient processing of 16S rRNA in E. coli (Bylund et al., 1998). The cca gene (tRNA aminoacylation) which encodes the enzyme Ccase (EC 2.7.7.72) that adds and repairs the 3’-terminal CCA sequence in tRNAs is not included in JCVI-syn3.0. In our results this gene is present to compensate the tendency to have tRNA without the CCA end in OS genomes (data not shown). Finally, nine ribosomal proteins that are present in the JCVI-syn3.0 genome are not include in our first proposal: rplJ (L10), rpmC (L29), rpmD (L30), rpmE (L31), rpmF (L32), rpmJ (L36), rpsP (S16), rpsR (S18), and rpsU (S21). It is known that the composition of the large subunit of the ribosome is less conserved than the one of the small subunit. Moreover, genes of S21, L30 and L31 ribosomal proteins have been consistently reported to be missing (Nikolaeva et al., 2021). Based on these facts, we have only added rpsP and rpsR to our minimal translation machinery.

The final minimal translation machinery proposed in our study (Figure 4) is composed of 142 genes, 113 of which are shared with JCVI-syn3.0 (76.8%), while 112 and 87 (78.8 and 61.3%) are essential in E. coli K-12 MG1655 and B. subtilis 168, respectively.