Towards functional characterization of archaeal genomic dark matter

The drop of sequencing costs over the last decade has led to a dramatic increase in the influx of new genomes into public databases. Furthermore, unlike the preceding years, most of these new genomic sequences are coming from metagenomic projects and belong to unculturable species [1,2]. In particular, metagenomics has yielded more than 10 major new archaeal groups including most of the lineages in the DPANN (Diapherotrites, Parvarchaeota, Aenigmarchaeota, Nanoarchaeota and Nanohaloarchaeota) superphylum that include, mostly, unculturable archaea with small genomes, many if not most of them, symbionts or parasites of other archaea. Affiliated with the TACK (Thaumarchaea, Aigarchaea, Crenarchaea, Korarchaea) superphylum is the Asgard group that currently consists exclusively of uncultured organisms of the putative phyla Loki-, Thor-, Odin- and Heimdallarchaeota lineages [3] that appear to be the closest archaeal relatives of eukaryotes. Additionally, new putative phyla have been discovered within the TACK superphylum including Bathyarchaeota, Geoarchaeota and Verstraetearchaeota, as well as among Euryarchaeota (Altiarchaea, Thalassoarchaea, Theionarchaea, Methanonatronarchaeia, Hadesarchaeota, Methanofastidiosa) [2,4].

For the uncultured archaea (and bacteria), gene annotation, based on comparison of protein sequences and operon organizations, is the only available source of information. Most often, the genome annotation that is deposited in public databases, such as GenBank, is generated automatically and thus is notably error-prone. Typical genome annotation errors include inaccurate gene calling whereby small ORFs (open reading frames) are falsely predicted as protein-coding genes, prediction of genes in the wrong DNA strand, prediction of ORFs in sequences that are actually non-coding, such as CRISPR arrays, and most often, erroneous assignment of start codons [5]. These annotation errors commingle with multiple gene fragments and pseudogenes that emerge through natural processes of gene degeneration. Arguably, most important, the gene annotation pipelines have to operate in a ‘safe mode', to minimize the rate of false-positive assignments. As a result, numerous sequences in the ‘twilight zone' of sequence similarity are annotated as ‘hypothetical proteins', a problem that affects, primarily, fast evolving genes, in particular, those involved in anti-parasite defense and genes that encode small proteins [6]. The quality of the annotation depends not only on the quality of the computational analyses themselves but also on the speed and completeness of the integration of new experimental data on protein functions integrated into annotation pipelines. At least in the case of archaea, annotation in public databases often runs behind experimental studies. For example, archaea-specific ribosomal proteins L45 and L47, experimentally identified in 2011 [7] and pre-rRNA processing and ribosome biogenesis proteins of the NOL1/NOP2/fmu family characterized in 1998 [8], are still not included in the annotation pipelines, so that most of the respective proteins remain ‘hypothetical'. The situation becomes even worse when it comes to the numerous confident predictions of protein functions that come from in silico analyses and cover several hundred protein families and several thousand ‘hypothetical' proteins. Among these predictions, there are several conserved families of membrane proteins [9], numerous genes linked to type IV pili systems [10], genes associated with various signal transduction pathways [11,12], polymorphic toxin systems [13], as well as integrated viruses and plasmids [14,15].

The mostly technical issues outlined above appear to stand behind the highest fractions of ‘dark matter' in microbial genomes. When more sensitive methods for sequence comparison and manual curation are employed, the ‘dark matter' fraction can be brought down to ∼20%, on average [16]. However, even after these substantial improvements in gene annotation, the dark matter includes millions of seemingly unique, completely uncharacterized proteins. Obviously, this number will grow fast as more genomes are sequenced, even if the fraction of dark matter in genomes remains constant or slowly drops.

An intriguing question of obvious importance is: what are the functions of these enigmatic genes? Comparative genomics and in-depth sequence analysis remain major approaches for prediction of protein functions, and the importance of this analysis further increases with the rising contribution of uncultured organisms to the genomic databases [17–19]. The growing volume and diversity of sequence databases is both a challenge — because of the increasing computational costs — and a boon to functional annotation of genes because the sensitivity of sequence searches can be dramatically increased thanks to the use of protein family profiles as queries. In addition to the increased sensitivity of sequence similarity detection, functional prediction strongly benefits also from comparative analysis of genomic contexts in increasingly diverse microbial genomes [17,20,21]. To reflect this, the COMBREX database for documenting experimental and in silico evidence for protein function predictions and for the prioritization of uncharacterized proteins for experimental testing has been developed [22]. Furthermore, experimental platforms for systematic validation of functional predictions produced by computational genomics have been launched, including characterization of new enzymes [23,24] and defense systems [25].

Several years ago, we undertook an initial analysis of dark matter islands in 168 archaeal genomes [26]. We found that these islands comprised ∼20% of archaeal genomes and that in-depth analysis allowed us to predict at least a general function for many of such loci and individual genes. In particular, it has been found that dark matter islands are enriched in integrated elements, novel defense systems and other genes implicated in interspecies conflicts [26]. Here, we report an analysis of the dark matter in 524 archaeal genomes covering all major archaeal lineages.

In the archaeal genomes from the GenBank version of March 2018, the fraction of dark matter genes (those annotated as ‘hypothetical' or ‘uncharacterized’ proteins) lies with the broad 30–80% range. The database of Archaeal Clusters of Orthologous Genes (arCOGs) has been employed to assign more annotations to archaeal genes [16]. When the arCOGs are used to annotate this set of genomes, the dark matter fraction (genes that do not belong to arCOGs or belong to the uncharacterized arCOGs of the functional category S) falls to 15–40% (Supplementary Table S1). Additionally, 8% of the genes assigned to arCOGs represent a ‘gray matter', i.e. arCOGs with a general function prediction only (functional category R, Figure 1A). Overall, the dark matter dominates the diversity of the gene families (92%) but represents a minority of genes (22%; Figure 1A). This difference stems from the fact that the dark matter families are typically small and are represented in only a few genomes (Figure 1B). Only 0.1% of the dark matter families are present in over 200 (out of the 524) genomes; for comparison, 22% of the functionally annotated arCOGs cross this threshold.

In most of the archaea, the number of the dark matter genes scales super-linearly with the genome size (Figure 1C) but two groups stand out in having inordinately high fractions of unannotated genes, DPANN and Thaumarchaea. The Asgard group contributes two more extreme outliers with much more dark matter than expected from their genome size (Lokiarchaeota archaeon CR_4 and Candidatus Heimdallarchaeota archaeon LC_3).

The protein sequences in the dark matter clusters tend to be much shorter compared with those that are functionally characterized (Figure 1D). The dark matter clusters have the median length of 93 amino acids, compared with 239 and 221 amino acids, respectively, for the ‘gray' and ‘bright' matter clusters. The small size of numerous dark matter genes is likely to result from a combination of at least five factors: (1) some of these ORFs are spurious and do not correspond to actual proteins, (2) the sensitivity of sequence similarity search drops with the protein size, so that short proteins are more likely to end up as dark matter, (3) proteins associated with certain functions, in particular, virus-encoded proteins, tend to be particularly small, (4) small proteins with a narrow phyletic distribution are less likely to attract attention of researchers, and therefore, are, in general, poorly characterized [6], (5) fast evolving lineages such as DPANN tend to have smaller ORFs.

Despite the lack of functional annotation, dark matter genes are far from being a completely random assemblage. This non-randomness can be gleaned from their spatial distribution in archaeal genomes. The 285 155 archaeal dark matter genes form 192 245 ‘islands' (contiguous blocks) in the 524 genomes, with the length of these islands distributed according to a power law-like heavy-tailed distribution (Figure 1E). The longest ‘dark matter island' consists of 49 genes and 205 islands (0.1%) are longer than 10 genes. In contrast, repeated sampling of 285 155 random genes leads to an exponentially declining the distribution of island lengths, with none exceeding 10 genes (frequency of <0.00005%).

An important minority of the dark matter genes encodes uncharacterized conserved proteins. Among the 218 genes in the pan-archaeal core, only one remains uncharacterized (arCOG04076, DUF359 protein family). Analysis of domain fusions suggests that proteins of this family are involved in CoA biosynthesis. However, there are many more functionally uncharacterized genes, with a broad distribution among archaea and often archaea-specific, that have been assigned to the common ancestor of all extant archaea, with a greater than 90% posterior probability [27]. These genes are expected to be involved in essential cellular processes and are prime targets for experimental study (Table 1). Although the structures of many of these proteins have been solved, only a few of them, in addition to DUF359, could be linked to a known pathway or cellular system, based on domain fusions, context analysis or similarity searches.

Asgard, DPANN and Thaumarchaea are especially rich in ‘dark matter' genes, which is not surprising because these are deep-branching, poorly characterized and mostly uncultured archaeal groups (Figure 1C). Despite this dark matter enrichment, there are only a few uncharacterized phylum-specific gene in Asgard and DPANN. In Asgard archaea (eight sequenced genomes), ∼500 arCOGs are represented in seven or eight genomes, and only three among these could not be assigned to arCOGs of the 2014 version [16] (Table 1). One of these three new arCOGs is the Vps25 subunit of the ESCRT-II complex that is implicated in cell division and/or membrane remodeling. The other was initially ‘uncharacterized', but HHpred search shows that one of these is a distant homolog of the eukaryotic signature protein gelsolin, an actin-binding protein. The Asgard archaea encode multiple gelsolin paralogs, but the proteins of this particular cluster were annotated as ‘hypothetical' because of extreme sequence divergence [3]. The paucity of Asgard-specific genes appears surprising because it could be expected that many eukaryotic signature genes found in these genomes would be ancestral and present in most or all Asgard genomes. This is, apparently, not the case although the Asgard genomes are still in the draft stage, so that some genes are likely to be missing.

The paucity of genes specific to the DPANN group (only 47 gene clusters are present in at least 17 out of 68 DPANN genomes and nowhere else) is less puzzling because most of these genomes are streamlined and lack many genes from the archaeal core. For two of the DPANN-specific gene clusters, cls.004259 and cls.004634, HHpred searches reveal similarity to minimal nucleotidyltransferase (MNT) and HEPN domains, and PepSY domain, respectively (Table 1), so these genes can be more appropriately classified as ‘gray', with a general functional prediction. The cls.004259 cluster, most probably, is a toxin–antitoxin module [28,29]. However, the unusual conservation of this gene, compared with the patchy distribution typical of most toxin–antitoxin systems, suggests that this protein plays some important role in the DPANN archaea. The YpeB, or double PepSY-like domain-containing proteins of the cls.004634 is likely to be an inhibitor of proteases that remain to be identified [30,31]. Two other DPANN-specific protein families remain enigmatic (Table 1).

In contrast, there are 53 Thaumarchaea-specific, functionally uncharacterized arCOGs that are represented in at least 90% of the thaumarchaeal genomes and absent from other archaea (Table 1). At present, sequence and genomic context analysis are not highly productive in the elucidation of the likely functions of these genes but they, obviously, are an important resource for experimental study.

Evidence is accumulating that processes of recombination and HGT in prokaryotes are not random [32,33]. Rather, these processes lead to the formation of islands of genes that are linked by common functional themes and/or evolutionary themes. Such islands include clusters or superoperons of house-keeping genes or superoperons [34], defense islands [35,36], islands of integrated elements [37–39], polymorphic toxins [13], virulence islands [40], and others [41]. The formation of genomic islands is driven, in part, by the selective advantages of the spatial clustering of functionally connected genes, such as the possibility of co-regulation, and in part, by non-adaptive ‘preferential attachment’ of non-essential genes, such as defense systems and mobilome components.

Analysis of genomic islands allows us to move many dark matter genes to the ‘gray' zone but often provides even for precise functional predictions. In particular, certain families of house-keeping protein families that evolve fast are retained in stable genomic contexts across long spans of genome evolution. Such a mismatch between sequence and positional conservation has been demonstrated, for example, for DNA replication initiation complex subunits of the GINS and Cdc48 families [42] and for the membrane insertase YidC [9]. Defense islands often include many dark matter genes, and for some of these, involvement in defense functions has been experimentally demonstrated or predicted by sequence analysis [25,36]. Numerous genomic islands that are rich in dark matter actually are integrated mobile genetic elements, such as casposons, proviruses and plasmids in archaea, and in many cases, their precise or approximate boundaries can be identified [14,15,26,43]. Fast evolving multigene systems often contain signature genes that are relatively well conserved, so that the location of such genes can point to the functions of the surrounding dark matter genes. Notable cases in point are CRISPR–Cas systems, with cas1 genes as a signature [44], viruses with capsid proteins as a signature [45], polymorphic toxins with Zn-dependent proteases of the DUF4157 family as a signature [13], and many others.

Figure 2 shows five examples of diverse islands in archaeal genomes containing multiple uncharacterized genes for which at least general functional predictions were feasible. Archaeal Type IV pili systems have been explored in detail, uncovering enormous diversity and signs of fast evolution, especially, in the case of pilins [10]. Therefore, it is not surprising that, in some of the DPANN genomes, predicted components of these systems do not show detectable similarity to the respective components from other archaea. Nevertheless, the presence of previously described components of Type IV pili systems, such as FliI, TadC and specific surface proteins, and the presence of signal peptides in the dark matter proteins suggest that all these proteins are diverse pilins (Figure 2A). Colicin D is one of the widespread toxins found in many polymorphic toxin systems in bacteria, often as a C-terminal domain that is fused to other domains involved in the toxin delivery [13]. Thus, the loci shown in Figure 2B, most probably, represent polymorphic toxin systems. The presence of PD-DExK nuclease domains, which also are abundant toxins in these systems, supports this conclusion. Multiple defense islands in archaeal genomes have been thoroughly studied because they contain CRISPR–Cas system components [46]. Many of the remaining ones include multiple TA systems consisting of two small genes that form a two-gene operon [29]. For such operons, if either toxin or antitoxin is known, it is most likely that the other, unannotated small gene in the operon is the respective counterpart (Figure 2C). Other genes present in these loci could be other, yet uncharacterized defense systems. The integrated MGE shown in Figure 2D corresponds to three distinct groups of viruses as indicated by the presence of the respective signature, namely, Pleolipoviridae (His2 major capsid protein), Caudovirales (terminase small and large subunits, TerS and TerL), Fuselloviridae (AAA ATPase, arCOG07960) [45]. All these elements contain numerous uncharacterized genes, supposedly, virion components, genes involved in viral replication, multiple inhibitors of host defense systems, etc. Even if prediction of specific function for most of these genes is currently out of reach, most of them can be confidently annotated as virus-related genes. The final example (Figure 2E) could represent a bacteriocin-like toxin or a quorum-sensing system. One of the genes encoded in this locus encodes a family C39 peptidase involved in bacteriocin precursor peptide processing [47]. Many of the other proteins encoded in this locus contain a leader peptide terminated by a double-glycine motif which is the characteristic recognition substrate of the peptidase [47]. Several genes in this locus are duplicated which is typical of systems involved in interspecies conflicts [13].

The genomic dark matter of archaeal and bacterial genomes presents both challenges and opportunities for research in microbial biology. Given that the fraction of dark matter remains (nearly) constant as new genomes are sequenced, the total amount and diversity of the dark matter increases rapidly with the growth of the genome database. Thus, there are more and more uncharacterized genes and also a greater capacity to infer their functions using increasingly efficient methods for sequence and genomic context analysis. To make the study of the dark matter informative and productive, carefully curated databases of gene families and improved transfer of annotations are essential. The computational analyses set the stage for systematic experimental investigation. Concerted effort on the functional characterization of the dark matter is likely to bring major pay-offs through improved understanding of poorly studied but crucially important aspects of microbial biology, primarily, various types of intergenomic conflicts and host–parasite coevolution. These processes are especially poorly understood in archaea, making the study of the dark matter particularly pertinent. Moreover, there is potential for the discovery of new defense systems that could be subsequently adopted as genome engineering tools as amply demonstrated by the discovery of different variants of CRISPR–Cas systems.

arCOGs

archaeal clusters of orthologous genes

DPANN

Diapherotrites, Parvarchaeota, Aenigmarchaeota, Nanoarchaeota and Nanohaloarchaeota

HTH

helix turn helix

MNT

minimal nucleotidyltransferase

ORFs

open reading frames

RHH

ribbon–helix–helix

TACK

Thaumarchaea, Aigarchaea, Crenarchaea, Korarchaea

K.S.M., Y.I.W. and E.V.K. are supported by intramural funds of the US Department of Health and Human Services (to National Library of Medicine).

The Authors declare that there are no competing interests associated with the manuscript.