Phylogenetic profiling for probing the modular architecture of the human genome

Abstract

Information about functional connections between genes can be derived from patterns of coupled loss of their homologs across multiple species. This comparative approach, termed phylogenetic profiling, has been successfully used to infer genetic interactions in bacteria and eukaryotes. Rapid progress in sequencing eukaryotic species has enabled the recent phylogenetic profiling of the human genome, resulting in systematic functional predictions for uncharacterized human genes. Importantly, groups of co-evolving genes reveal widespread modularity in the underlying genetic network, facilitating experimental analyses in human cells as well as comparative studies of conserved functional modules across species. This strategy is particularly successful in identifying novel metabolic proteins and components of multi-protein complexes. The targeted sequencing of additional key eukaryotes and the incorporation of improved methods to generate and compare phylogenetic profiles will further boost the predictive power and utility of this evolutionary approach to the functional analysis of gene interaction networks.

Significant similarity between two DNA or amino acid sequences is used to infer shared ancestry, or homology, of the DNA elements or proteins being compared. A high degree of sequence similarity between homologs strongly indicates a conserved biological function, a cornerstone of comparative genomics that has been used to provisionally assign functions to thousands of human genes based on decades of detailed experiments in vertebrate and invertebrate model systems. However, the differences in sequence between (or the complete loss of) homologs evolving independently in separate lineages encode information as well: a close functional coupling between unrelated genes (or non-coding genetic elements) often manifests itself in correlated patterns of sequence similarity across species (de Juan et al., 2013), a fact that can be exploited to discover novel functional links. Specifically, the inference of functional connections between protein-coding genes based on shared binary patterns of homolog presence and loss is termed phylogenetic profiling (Figure 1A) (Pellegrini et al., 1999).

Figure 1. Phylogenetic profiling and its challenges.

(A) This schematic illustrates the method of phylogenetic profiling using homologs (orthologs) mapped across 8 species that are related through the accompanying species tree. Protein 1 and 2 interact functionally and share identical homolog (ortholog) distributions, with each potential ortholog represented in a binary phylogenetic profile with a 1 (black) if present or a 0 (white) if absent. Inferred gene loss events are highlighted on the tree. The correlated phylogenetic profiles for Proteins 1 and 2 can be used to predict the conserved functional interaction. (B) Challenges to phylogenetic profiling. In this illustration, an ancestral protein duplicates once (box inset), leading to a complex distribution of orthologs in different extant species. The graph below the species tree represents a theoretical distribution of similarity scores (e.g. BLAST) generated against human protein 1. Each point (a putative ortholog in each species) is color-coded in accordance with its reciprocal best match in the human genome (BBH; best bidirectional hit). Species branching off before the duplication event contain only one homolog resulting in an artifact-prone BBH match (red arrows). The dotted line represents a suitable homology threshold. Phylogenetic profiles are generated using either this homology threshold (top) or a best bidirectional hit criterion (BBH, bottom) in each species. Errors in the phylogenetic profiles resulting from each method are highlighted with red dotted lines.

Phylogenetic profiling exploits a specific evolutionary scenario, namely one in which a pair or larger group of genes are functionally coupled in such a way that the loss of one component leads directly or indirectly to the loss of the others (Figure 1A). While this scenario can only apply to a subset of all possible genetic interactions, a close correlation between binary phylogenetic profiles is frequently associated with them being part of the same physical protein complex, metabolic cascade or regulatory module (Pellegrini, 2012), providing a powerful approach to predict functions for unknown genes and define interdependent genetic modules. This perspective focuses primarily on novel functional insights that can be gained by phylogenetic profiling applied to the human genome (for a general overview of comparative genomics, see Alföldi and Lindblad-Toh, 2013).

Phylogenetic profiling already showed much promise as a predictive tool during its first application to bacterial gene sets just before the end of the millennium (Pellegrini et al., 1999), but the first fully sequenced genome for a multicellular eukaryote had only just been released (The C. elegans Sequencing Consortium, 1998). The 15 years since have seen an unprecedented increase in the number of eukaryotic genomes driven by plummeting sequencing costs. This led to the successful application of phylogenetic profiling to a genome-wide analysis of S. cerevisiae (Marcotte et al., 1999), the discovery of novel Drosophila cilia genes (Avidor-Reiss et al., 2004), a screen for novel small RNA pathway components in C. elegans (Tabach et al., 2013a), and multiple components of a key mitochondrial uniporter (Baughman et al., 2011; De Stefani et al., 2011). In recognition of the method’s utility, web servers for comprehensive phylogenetic profiling continue to be developed (Cheng and Perocchi, 2015) and coevolution metrics have been incorporated in some major interactome databases (von Mering et al., 2005; Szklarczyk et al., 2015).

Three recent studies have systematically investigated the utility of phylogenetic profiling in revealing genetic interactions between human genes (Dey et al., 2015; Li et al., 2014; Tabach et al., 2013b). Tabach et al. mapped hundreds of co-evolving human gene sets (identified using correlated homology scores) and disease annotations (Tabach et al., 2013b), a valuable dataset subsequently utilized to identify novel components of the mammalian meiotic methylation program (Schwartz et al., 2013). Li et al. used statistical inference to expand groups of correlated human phylogenetic profiles into larger modules, generating predictions for approximately 150 cellular pathways and complexes (Li et al., 2014). A recent approach taken by our group extended phylogenetic profiling to ‘orthogroups’ of homologous human genes and calculated a genome-wide matrix of all pairwise co-evolution scores, identifying a much larger set of modules (Dey et al., 2015). Experiments in our study as well as subsequent studies have validated a subset of functional predictions related to primary cilium function and novel interactors of the WASH complex (Phillips-Krawczak et al., 2015). The success of these studies in driving empirical discovery is of particular relevance to biomedical science given the large proportion of the human protein-coding genome that remains poorly characterized (Dey et al., 2015).

This article focuses on how to build on these recent successes and effectively leverage the growing pool of available genome sequences. We argue that sequencing more free-living protists is a vital step in the accurate reconstruction of eukaryotic gene histories. We discuss a role for phylogenetic profiling in the investigation of human cellular function through comparative biology. Finally, we examine the modular architecture retained for some, but not all, cellular processes across diverse ecological and cellular niches through millions of years of eukaryotic evolution.

Optimizing predictive phylogenetic profiling

Use of hierarchical groups of orthologs.

Generating a phylogenetic profile for a human gene involves first identifying its orthologs in other species (homologs derived vertically from a common ancestor and expected to share the same function; Koonin, 2005). Orthology inference is a mature field, with a large number of graph-based (clustering based on sequence similarity scores, e.g. BLAST) and tree-based (reconciliation of gene trees, inferred from sequence similarity, with the species tree) algorithms (Huerta-Cepas et al., 2014; Li et al., 2003; Powell et al., 2014; Schreiber et al., 2014; Tatusov et al., 1997; Vilella et al., 2009). Even straightforward graph-based methods like the best bidirectional hit (BBH; orthology is assigned if the top-scoring homolog in a second species returns the original query gene in a reciprocal similarity search) sometimes outperform more complex tree-based approaches in comparative analyses (Kristensen et al., 2011; Trachana et al., 2011). Moreover, it should be noted that incomplete genome annotation and low homology scores at large evolutionary distances generate algorithm-independent errors; the latter can be partially addressed by using sensitive search methods like PSI-BLAST (Altschul, 1997) or delta-BLAST (Boratyn et al., 2012) that leverage additional information derived from conserved domains or secondary structure.

The scalability and easy implementation of graph-based approaches make them attractive for phylogenetic profiling, with some studies directly using homology thresholds (Li et al., 2014). However, even a single gene duplication can introduce a conceptual challenge: now, some species only carry a single gene with homology to two separate human genes (Figure 1B). Each time a gene is duplicated, the daughter genes, now capable of evolving independently, can diverge by acquiring new functions (neofunctionalization) or sharing the function of the parent (subfunctionalization) (Conant and Wagner, 2003; Conant and Wolfe, 2008). Thus, neither daughter gene is (by itself) a true functional ortholog of the non-duplicated gene found in lineages that branched off prior to the duplication event. Problematically, using homology thresholds will generate near-identical phylogenetic profiles for both daughter genes despite their possible functional independence (Figure 1B), and the BBH criterion can cause mismatches in species that branched off before the duplication event (Figure 1B) (Dalquen and Dessimoz, 2013).

While this challenge can be circumvented by eliminating all human genes with detectable human homologs (co-orthologs) from the analyzed set (Li et al., 2014; Tabach et al., 2013b), this represents only a partial solution because an overwhelming fraction of human genes are derived from historical duplication events (Cotton and Page, 2005; Dey et al., 2015). First, the vertebrate lineage carries clear signatures of two genome-wide duplications (Blomme et al., 2006). Second, many human gene families of fundamental importance to cell biology have a demonstrated history of broad expansion coupled with functional divergence (Gu et al., 2002; Lespinet et al., 2002): GPCRs (Bjarnadóttir et al., 2006), small GTPases (Boureux et al., 2007), and kinases (Shiu and Li, 2004), to name just a few.

A more inclusive solution is to sequentially group co-orthologs in the same genome into ‘orthogroups’ (Figure 2A). Each orthogroup represents the extent of sequence space (and implied functionality) that the daughter genes have explored after duplication. Other genomes can then be queried for a reciprocal match to any of the coorthologs within the group (Figure 2A). Consequently, methods that generate a separate phylogenetic profile for each orthogroup (Dey et al., 2015; Wapinski et al., 2007) enable a comprehensive exploration of the functional prediction space without excluding gene families from analysis.

Figure 2. Improvements to phylogenetic profiling.

(A) The schematic illustrates the use of an orthogroup strategy to resolve a single gene duplication event. First, phylogenetic profiles are generated for Daughter 1 and Daughter 2 using a BBH criterion in species that branched off after the duplication event. Next, an orthogroup (group of co-orthologs or sister genes) is created that contains the two human proteins Daughter 1 and Daughter 2. A third phylogenetic profile can now be generated that assigns an ortholog to a species that contains a BBH match to either Daughter 1 or Daughter 2. (B) Schematic to illustrate the strengths and weaknesses of common algorithms used to quantify the strength of coevolution between phylogenetic profiles using two different evolutionary scenarios (Case 1 and 2) for Gene 1 and Gene 2. In both scenarios, phylogenetic profiles for Gene 1 and Gene 2 are being compared across 18 species related through the accompanying species tree. The total number of shared profile presence calls (13) and absence calls (5) are identical in each case. Case 2 contains more inferred loss events (red stars) and shared transitions between phylogenetic profiles (red wedges). Blue text indicates the relative strength of coevolution assessed by linear, model-based and runs-based algorithms for these two scenarios.

Optimizable measures of coevolution.

In principle, since independent losses in multiple lineages are good indicators of functional coevolution (Figure 2B; case 2 represents a higher likelihood of functional co-evolution than case 1), the most rigorous way to compare phylogenetic profiles involves modeling gene gains and losses on each branch of the complete species tree. Parsimony and maximum likelihood methods have been used successfully in the past for small numbers of bacterial and fungal genomes (Barker and Pagel, 2005; Barker et al., 2007). Most recently, Li et al. developed an algorithm to generate statistical models for gene gain and loss from pre-selected seed groups already annotated to be part of the same pathway, and search the human genome for additional genes conforming to the model (Li et al., 2014). Though statistically rigorous, their approach relies on pre-existing pathway annotations and is insensitive to co-evolution at the scale of individual gene-pairs, making it unsuitable in its current form for an unbiased genome-wide analysis in humans.

The alternative is to use a heuristic score, which comes with the advantages of rapid optimization against functional interaction resources and the ability to scale with both genome complexity and the number of genomes. Unfortunately correlation scores that give each species equal weight produce artifacts, as a single gene loss event can result in drastically different ortholog distributions depending on where it occurs within the tree (Figure 2B) (Kensche et al., 2008). This effect can be partially neutralized by sampling an even distribution of species (Tabach et al., 2013a), though with the caveat of assuming a uniform probability of gene gain/loss across lineages that encounter widely varying ecological niches and selective pressures.

One effective strategy that combines the strengths of both approaches listed above involves using shared ‘runs’ (Cokus et al., 2007) or transitions (Dey et al., 2015) in phylogenetic profiles to indicate independent loss events (Figure 2B). These scoring schemes incorporate information from the species tree without requiring full models of gain and loss, making them easy to optimize and scale up to thousands of genes across hundreds of species. Drawing inspiration from tree-based methods, further heuristic constraints derived from evolutionary logic and parsimony (penalties for unlikely losses and down-weighting the influence of parasite genomes, for example) could reduce false positive rates and increase the sensitivity of predictions.

Leveraging eukaryotic diversity

As more and more species get sequenced, it is increasingly clear that almost a quarter of human genes can be traced to the earliest eukaryotes (Koonin, 2010) and have since been lost in many plant, fungal and parasitic protist lineages. This number was initially underestimated, largely because many supposedly early-branching species in a “crown-group” model of the eukaryotic tree were assigned erroneous positions caused by fast rates of genome evolution and parasitic lifestyles (Stiller and Hall, 1999). Far from being ‘primitive’ pre-mitochondrial organisms, parasites such as Giardia lamblia actually represent the results of reductive evolution from a complex ancestor that possessed fully functional mitochondria (Embley and Martin, 2006). In contrast, the genome of the recently sequenced free-living Naegleria gruberi (Fritz-Laylin et al., 2010, 2011) is much closer to that ancestral state, encoding complete actin and microtubule skeletons, complex transcriptional and signaling machinery (including GPCR, histidine kinase modules and twice as many adenylate/guanylate cyclases as humans), as well as thousands more spliceosomal introns than its parasitic relative Trypanosoma brucei (Siegel et al., 2010).

The unanticipated degree of conservation of ancient eukaryotic machines revealed by these analyses opens up new possibilities for systematic comparative biology (Box 1). Importantly, the many distinct lineages (Burki, 2014) of unicellular protists represent a huge reservoir of genomic diversity that can play a major role in informing phylogenetic profiles. Figure 3 illustrates this argument by highlighting the overall contribution of ortholog losses/absences in individual species to the representative phylogenetic profiles of evolutionary modules of human genes (Figure 3A and 3B) (Dey et al., 2015). The protists exhibit a two-fold dynamic range of shared gene content: some species such as the free-living Naegleria gruberi have more orthologs for human genes than any fungi or plants while others, particularly parasites, have undergone severe reductive evolution (Figure 3C). This diversity in gene loss greatly increases the likelihood of observing informative and unique phylogenetic profiles, explaining why losses in protists contribute to over a third of all coevolving modules identified (Figure 3B).

Box 1. Evolutionary cell biology.

Experimental cell biology informed by genomics and evolutionary theory (evolutionary cell biology) has the potential to provide powerful new insights into complex cellular functions (Lynch et al., 2014). From the human perspective, this could involve comparing and contrasting function of a conserved module, regulatory motif or protein across multiple species, but also the parallel investigation of a complex intracellular feature in a potentially reduced or more ancestral environment.

A striking example is provided by the flagellum of the green alga Chlamydomonas reinhardtii, a structure that bears striking structural and functional similarities to the mammalian cilium (Kozminski et al., 1993; Rosenbaum and Witman, 2002). Enabled by the ease of laboratory culture, classical genetics and the generation of non-lethal flagellar mutations, experiments in Chlamydomonas have resulted in deep insights into ciliary function in mammalian health and disease, and helped define a minimal set of genes required for cilium function dating back to the earliest eukaryotesall this from a unicellular alga separated from humans by 10^9 years of evolution (Li et al., 2004; Pazour et al., 2005; Silflow and Lefebvre, 2001). Recent phylogenetic profiling studies have only served to further reinforce the extent of the functional coupling of cilia components across these two species (Avidor-Reiss et al., 2004; Carvalho-Santos et al., 2011; Dey et al., 2015; Li et al., 2014).

The case of the cilium raises the question of how such an evolutionary perspective can be systematically extended to other cellular modules. The advent of adaptable genome-editing technology, super-resolution imaging, and the abundance of high quality genome data make it possible to generalize this approach to species drawn from across the eukaryotic tree and other domains of life, developing novel experimental systems on reasonable timescales. In doing so, we might choose to complement studies in vertebrates and mammals by investigating splicing in Giardia lamblia (Nixon et al., 2002), sperm development in Nematostella vectensis (Putnam et al., 2007), GPCR signaling in Naegleria gruberi (Fritz-Laylin et al., 2010), and the origins of the nervous system in comb jellyfish (Moroz et al., 2014). Phylogenetic profiling can help to guide this approach by providing a comprehensive map of orthologs across species but also a set of core functional interactions that have been preserved across evolutionary timescales.

Figure 3. Species contributing to informative phylogenetic profiles.

(A) Schematic illustrating the generation of co-evolving modules. Phylogenetic profiles of all human genes and orthogroups (Dey et al., 2015) were clustered using an agglomerative algorithm to generate 334 modules containing 3 or more components (illustrated using two modules containing 3 and 4 components respectively with phylogenetic profiles in gray). An averaged binary phylogenetic profile was generated for each module (black).

(B) Averaged module phylogenetic profiles spanning 177 eukaryotic species ordered by 13 major branches, represented using text labels as well a color bar (the complete list of species can be found in Figure S1 of Dey et al., 2015). Each binary profile represents the consensus (averaged, 1:black, 0:white) for each one of the 334 strongly coevolving modules. (C) The fraction of modules missing or lost in each species, estimated from (A). Each point represents 1–(sum of column in (A)), color coded in accordance with the species tree. Groups of species populating the same branch with a mixture of low and high loss fractions (big spread on y-axis) contribute strongly to the identification of coevolving modules through informative lineage-specific gene losses.

We have a long way to go: approximately 80% of all existing or ongoing eukaryotic genome projects are restricted to opisthokonts (fungi, animals and choanoflagellates) and multicellular plants, with very little coverage of the other major, diverse eukaryotic supergroups (Dawson and Fritz-Laylin, 2009). Of the few protists that have been sequenced, most are parasites. This clear shortfall has led to a call by some groups for a concerted effort to sequence more aquatic free-living protists (Dawson and Fritz-Laylin, 2009; Keeling et al., 2014), and we emphasize here the relevance of such projects for functional predictions in humans. Though not without its challenges (primarily the difficulty of growing many such species in pure laboratory culture), metagenomics anchored by high-quality reference genomes can help accelerate this process (Heywood et al., 2010).

Discovering evolutionary modularity

Evolutionary cohesion.

Biological networks are widely considered to be intrinsically modular, consisting of sub-networks isolated chemically or spatially from the rest of the network and carrying out discrete functions. However, there are many ways to partition complex systems, and distilling modules from cellular networks has been an important focus of research in many different fields for many years, including developmental biology (Bolker, 2000), metabolism (Jeong et al., 2000; Segrè et al., 2005), signaling (Atay and Skotheim, 2014; Bhattacharyya et al., 2006; Lauffenburger, 2000; Meyer and Teruel, 2003), evolutionary biology (Roth, 1991; Wagner, 1996) and bioengineering (Alon, 2007). Modules are often defined empirically (Tanay et al., 2004) - a set of protein interactions restricted to a single subcellular compartment, a transcriptional circuit only active during a specific developmental stage, or a set of metabolic enzymes linked through a linear chain of substrates and products.

However, developing computational strategies to infer functional modules has become a priority (Alon, 2007) with the advent of comprehensive interactome maps (Rolland et al., 2014). Drawing on principles of circuit design and real-world scale-free networks (Alon, 2007; Barabási and Oltvai, 2004), studies have focused on the identification of characteristic topological ‘motifs’ (Milo et al., 2002; Shen-Orr et al., 2002). While powerful, these methods can be confounded by the characteristic hierarchical organization of many regulatory features (Papin et al., 2004), and also simply by errors, incomplete coverage and the absence of dynamical measurements in most high-throughput data sets (Alexander et al., 2009).

In search of a complementary strategy, it is perhaps instructive to note that the very definition of modularity in some early papers was based on conservation of homologous structures across species (Roth, 1991; Wagner, 1996). Although there has been considerable debate over the specifics of how modularity itself might evolve (Espinosa-Soto and Wagner, 2010; Kashtan and Alon, 2005; Wagner et al., 2007; Wang and Zhang, 2007), it is nonetheless clear that genes can exhibit “evolutionary cohesion”-be gained and lost together in genomes encountering different environments (Campillos et al., 2006; Snel and Huynen, 2004)- and exert constraints on the evolution of their components (Chen and Dokholyan, 2006). Modules identified by phylogenetic profiling, a generalization of the cohesion principle, represent a functional coupling maintained across tissue, species and environmental context—an integrated ‘experiment’ across hundreds of experimental conditions impossible to recreate in the laboratory. It is notable that different approaches to human phylogenetic profiling converged on a highly overlapping set of functional modules distinct from those found using other methods (Dey et al., 2015; Li et al., 2014) and predicted novel sub-functions for protein complexes or pathways that merit empirical follow-up. For example, we identified a module containing FANCI, FANCD2 and FANCL that represent proteins belonging to two separate and well-characterized physical complexes involved in the Fanconi Anemia DNA damage sensing pathway (Figure 4A) (Dey et al., 2015; Moldovan and D’Andrea, 2009).

Figure 4. Evolutionary modularity.

(A) Left, schematic illustrating the components of two interacting physical complexes involved in the Fanconi Anemia pathway. Right, phylogenetic profiles for each protein in the two complexes 1:black, 0:white). Data from Dey et al. 2015. A red font is used to illustrate the coevolving phylogenetic module. Abbreviations for species branches: M: Mammals, V: Other Vertebrates, LT: Lancelets/Tunicates, EH: Echinoderms/Hemichordates, A: Arthropods, N: Nematodes, C: Cnidaria, SP: Sponges/Placozoa, CF: Choanoflagellates, F: Fungi, AB: Amoebozoa, P: Plantae, PR: Other protists. (B) Top, phylogenetic profiles for 6 proteins/ protein families involved in canonical EGFR signaling. Bottom, phylogenetic profiles for 6 enzymes involved in heme biosynthesis. Data from Dey et al. 2015. Species ordered as in (A).

Constraints on network evolution.

As highlighted earlier in this perspective, modules identified through phylogenetic profiling conform to an evolutionary model where the components of the module are interdependent and relatively isolated from the rest of the network: phylogenetic profiling results in functional predictions for approximately 10–15% of the human genome (Dey et al., 2015). Interestingly, however, in all three human profiling studies (Dey et al., 2015; Li et al., 2014; Tabach et al., 2013b), the highest scoring pairs or modules were enriched for metabolic, transport and structural functions, and depleted of canonical signaling proteins and regulators of transcription.

To illustrate this point with an example, we contrast the phylogenetic profiles of members of the EGFR signaling cascade (Figure 4B, top) with 6 enzymes involved in heme biosynthesis (Figure 4B, bottom). Similar trends have also been highlighted by bacterial phylogenetic profiling studies (Campillos et al., 2006), strongly suggesting the existence of generalizable constraints. Biological networks evolve through the gain and loss of nodes (gene duplication and loss) and the gain, loss and exchange of edges (new, lost or rewired functional links between proteins). However, not all edges are identical: there are more ways to alter (or generate) a kinase-substrate interaction or a transcription factor-binding site interaction than the specificity of an enzyme or binding interactions within a physical complex. This is reflected in the observation of pervasive rewiring in kinase and transcription factor interactions (Baker et al., 2012; Pearlman et al., 2011), at a faster rate than in metabolic and PPI networks (Shou et al., 2011). It might be expected that in these networks pervasive edge changes would lower the likelihood of modular gene loss occurring under selective pressure or following duplication events, leading to a depletion of signaling and transcriptional regulators from sets of correlated phylogenetic profiles. On the other hand, modular gene gain and loss might dominate the phylogenetic signal from stoichiometric physical complexes or metabolic cascades. These arguments support the existence of powerful constraints on network evolution revealed through the analysis of evolutionary modules that can be incorporated into future topological investigations of large-scale interactome maps.

Conclusion

Only a small fraction of the human genome encodes proteins, and major projects have been undertaken to investigate the function and evolution of non-coding regulatory elements on a genome-wide scale (Boyle et al., 2014; Gerstein et al., 2012). The massive scale of these projects has drawn attention away from the fact that the majority of protein-coding genes still remain completely or partially uncharacterized. With comparative genomics entering the mainstream and a 1000-dollar human genome becoming an imminent reality, there has never been a better time to leverage information from sequence coevolution to study human protein-coding genes and their interactions. In particular, phylogenetic profiling has the potential to mature as a powerful tool for human gene function discovery, especially with further technical refinements and the sequencing of key protist genomes. As highlighted in this article, despite its conceptual simplicity, analyzing the human genome through the lens of gene gain and loss has broad consequences for our understanding of eukaryotic genetic diversity, modular constraints on the evolution of genetic networks, and the capacity to drive evolutionary cell biology approaches to studying fundamental cellular functions across diverse experimental systems.