Computational Characterization of Moonlighting Proteins

Abstract

Moonlighting proteins perform multiple independent cellular functions within one polypeptide chain. Moonlighting proteins switch functions depending on various factors including the cell type in which they are expressed, cellular location, oligomerization status, and the binding of different ligands at different sites. Although an increasing number of moonlighting proteins have been experimentally identified in recent years, the quantity of known moonlighting proteins is insufficient to elucidate their overall landscape. Moreover, most moonlighting proteins have been identified as a serendipitous discovery. Hence, characterization of moonlighting proteins using bioinformatics approaches can have a significant impact on the overall understanding of protein function. In this work, we provide a short review of existing computational approaches for illuminating the functional diversity of moonlighting proteins.

1 INTRODUCTION

With the increase of the number of functionally well- characterized proteins as well as the advancement of large-scale proteomics studies, more and more proteins have been observed to exhibit more than one cellular function. These proteins were named as “moonlighting” proteins first by Jeffrey [1]. A moonlighting protein demonstrates multiple autonomous and usually unrelated functions. The diversity of dual functions of these proteins is in principle not a consequence of gene fusions, splice variants, multiple proteolytic fragments, homologous but non-identical proteins or varying post-transcriptional modifications. Moonlighting proteins are not limited to a certain type of organism or protein family, nor do they have common switching mechanisms through which they moonlight. The known mechanisms for switching functions include cell type of expression, cellular localization, oligomerization state, and identity of binding ligand [1].

Piatigorsky and Wistow [2] identified that crystallin, a structural protein in the eye lens of several species, also has enzymatic activity: this was one of the first examples of multifunctional proteins. Many known moonlighting proteins were originally recognized as enzymes, but there are also others that are known as receptors, channel proteins, chaperon proteins, ribosomal proteins, and scaffold proteins [1,3,4]. The secondary or moonlighting functions of these proteins include transcriptional regulation, receptor binding, involvement in apoptosis, and other regulatory functions. So far the identification of moonlighting proteins has been done by experiments and there exist comprehensive reviews of these proteins in literature [1,3–6]. Studies [7–9] suggest significant impacts of moonlighting proteins in diseases and disorders. Despite the potential abundance of moonlighting proteins in various genomes and their important roles in pathways and disease development, the number of currently confirmed moonlighting proteins is still too small to obtain a comprehensive picture of the cellular mechanisms underlying their functional diversity. This quantitative insufficiency is in large part due to the tendency for the additional function of these proteins to be found serendipitously in the course of unrelated experiments. Hence, a systematic bioinformatics approach could make substantial contributions in identifying novel moonlighting proteins and also in elucidating functional characteristics of moonlighting proteins.

In this article, we provide a short review of the existing computational analyses on moonlighting proteins. First, we discuss two works that investigated whether existing sequence-based function prediction methods can identify distinct dual functions of moonlighting proteins [10,11]. Then we review another work by Gomez et al. [12] on analysis of protein-protein interactions (PPIs) of moonlighting proteins where they studied whether the interacting partners of moonlighting proteins disclose the moonlighting function or not. Third, we analyze the work by Hernandez et al. where they explored structural aspects of known moonlighting proteins to identify whether the promiscuous functionality of these proteins are caused by the conformational fluctuations in their structures [13]. Then, we introduce recently developed databases of moonlighting proteins [14]. Lastly, we discuss the current situation of Gene Ontology (GO) [15] annotations of known moonlighting proteins in the UniProt database [16].

2.1 Moonlighting proteins pose a challenge in bioinformatics research

A review by Jeffery [4] discusses moonlighting proteins in the context of systems-level proteomics studies and presents challenges for computational analyses. Most sequence-based function prediction methods are based on homology searches or motif/domain identifications. Moonlighting proteins can complicate this approach since there are cases that orthologous proteins in different organisms do not share moonlighting functions. Moreover, the possibility of a moonlighting function would change how we treat the existence of motifs in the protein that have been identified with less confidence, since those hits may explain the moonlighting function of the protein. From a structural point of view, moonlighting proteins could be identified by discovery of multiple ligand binding sites. Last but not least, moonlighting functions have implications in the discovery of drug targets and biomarkers, since the knowledge of all functions of a target protein is necessary to design drugs that only affect the desired function of the target.

2.2 Sequence based function prediction on moonlighting proteins

Conventional sequence-based functional annotation methods are based on the concept of homology [17,18] or conserved motifs/domains [19–21]. Two works have investigated how well current sequence-based methods identify the distinct dual functions of moonlighting proteins. In one of the works, we have benchmarked performance of three sequence based function prediction methods, the Protein Function Prediction (PFP) algorithm [22,23], the Extended Similarity Group (ESG) algorithm [24], and PSI-BLAST [25], on a set of experimentally known moonlighting proteins [5]. PFP extends a traditional PSI-BLAST search by extracting and scoring GO annotations from distantly similar sequences and then applying contextual associations of GO terms observed in the annotation database. ESG performs an iterative sequence database search and assigns probabilities to GO terms. PFP and ESG have different characteristics: PFP is designed to have larger coverage by retrieving annotations from weakly similar sequences while ESG is for better specificity by taking consistently predicted GO terms from an iterative search.

In the performance evaluation for predicting the diverse functions of moonlighting proteins [5], we compared the predicted GO terms by PFP, ESG, and PSI-BLAST with those from both primary and moonlighting functions. In the average precision-recall for the 19 moonlighting proteins, ESG showed the highest precision for a recall range of 0.4–0.7 while PFP outperformed the other methods in recall. ESG had lowest recall among the three methods except for 5 cases. For PSI-BLAST, we used BLOSUM45 and BLOSUM30 in addition to the default BLOSUM60 in order to consider more distantly related sequences. Recall by PSI-BLAST improved using BLOSUM45. In the head-to-head comparison against PFP, PSI-BLAST with BLOSUM45 showed a higher recall than PFP for 8 proteins while PFP had a higher recall in 10 cases (1 protein had a tie). PSI-BLAST with BLOSUM30 failed to predict any GO terms above an E-value of 0.01 for 12 proteins. Overall, PFP and PSI-BLAST with BLOSUM45 showed higher recall than the rest of the methods.

These results highlighted the power of PFP in predicting the diverse functions of moonlighting proteins with high recall. Incorporating the BLOSUM45 matrix enhanced the power of PSI-BLAST greatly, which provides another indication that considering weakly similar sequences enhances the prediction of moonlighting functions of proteins.

The second work, by Gomez et al., compared the performance of homology-based and motif/domain-based methods in retrieving sequences with primary and/or moonlighting functions using a dataset of 46 moonlighting proteins [11]. They compared PSI-BLAST, and ten motif/domain-based methods.

For a dataset of 46 moonlighting proteins, the authors ran the eleven methods and retrieved all the sequences that matched a query protein above a certain standard score cutoff. If any of the retrieved sequences had the primary/secondary function of the query moonlighting protein, it was considered as a “positive match” for that corresponding function. For example, for the moonlighting protein FtsH (primary function: protease, moonlighting function: chaperone), the PSI-BLAST output contained two matched sequences (both with E-value 0.0): gi5231279 and gi12724524, which are a proteinase and a heat shock protein, respectively. Both sequences were considered a “positive match”, the first for the primary function and the second for the moonlighting function. Among the methods tested, PSI-BLAST outperformed others in finding positive matches for both functionalities of the moonlighting proteins. Among the ten motif/domain-based methods, PRODOM performed best. Among the 46 proteins in the dataset, the authors performed structural analysis on four proteins (BirA biotin synthetase, thymidine synthase, aconitase and fructose 1,6 biphosphatase) and found two different functional sites for three of them.

2.4 Exploring moonlighting proteins in protein-protein interaction networks

Protein-protein interaction (PPI) networks are a useful clue of protein function because proteins of the same biological function or pathways tend to interact [26–31]. Gomez et al. analyzed PPI networks of known moonlighting proteins to determine whether interacting proteins of moonlighting proteins possess the secondary functions of the moonlighting proteins [12].

A set of experimentally identified moonlighting proteins that have known interacting partners in the APID database were selected for this analysis [32]. Among these interacting partners, 605 proteins were selected that have GO annotations (in the Biological Process or Molecular Function categories) that match the function description of the moonlighting function of the query protein. For each of these selected interacting partners of a moonlighting protein, GO terms related to the moonlighting function were collected and a GO term enrichment score (p-value from the hypergeometric distribution) was computed using the GOStat package in R. Using a p-value cutoff of 0.05, the authors analyzed whether secondary functions of moonlighting proteins could be predicted. Among the six PPI databases they analyzed (MINT, DIP, BioGRID, IntAct, HPRD and BIND), DIP had the highest percentage of identifying the moonlighting function among its interacting partners (0.833) and MINT had the lowest percentage (0.6). The authors concluded that PPI networks contain information that discloses additional functionalities of moonlighting proteins.

2.5 Moonlighting Proteins and Disordered Regions

Intrinsically disordered regions have been found to have important roles in protein function [33]. The functional diversity of moonlighting proteins could be caused by structurally disordered regions as different conformations of disordered regions may facilitate different functions of a protein or allow a protein to interact with different protein partners.

Tompa et al. [34] reported earlier that some known moonlighting proteins have disordered regions, with which they bind the same partner in different conformations and at different binding sites, resulting in opposite effects of inhibiting or activating their interaction partners.

While some moonlighting proteins exhibit dual function due to disordered regions, this is not the case in the majority of moonlighting proteins. Hernandez et al. [13] investigated whether moonlighting proteins tend to have intrinsically disordered regions. 28 known moonlighting proteins were analyzed. Disordered regions of these proteins were predicted by four programs, PrDos, DisEMBL, Disopred, and IUpred. It turned out most of the moonlighting proteins do not have long disordered regions and are not considered as members of the Intrinsic Disordered Protein (IDP) class, which is defined as proteins that have more than 40 residues in disordered regions [35]. Most of the predicted disordered regions for these moonlighting proteins were quite short amino acid stretches, and in many cases were located at the N- or C-terminal regions of the proteins. Based on these results, the authors concluded that most moonlighting proteins do not fall into the IDP class.

2.6 Database of Moonlighting proteins

Currently, there exist three databases of moonlighting proteins. One of them, named MultitaskProtDB [14], has compiled 288 multitasking/moonlighting proteins at the time of this writing. This database lists known moonlighting proteins extracted from ten review articles. In addition, the authors performed text mining on articles in PubMed to identify moonlighting proteins using following keywords: moonlight/moonlighting proteins/enzymes, multitask/multitasking proteins/enzymes and gene sharing. The database holds 288 moonlighting proteins from ~100 difference organisms, among of which 91 are from human (32%), 23 from yeast (8%), 23 from Arabidopsis (8%) and 20 from E. coli (7%).

For each protein, users can retrieve its NCBI code, UniProt accession number, species information, canonical and moonlighting functions, PDB codes (if available), oligomeric state (if available), and reference to the corresponding literature. Interestingly, from the database the authors found that the most prevalent canonical/moonlighting GO pair is enzyme/nucleic acid binding proteins (74 out of 288): For example, proteins that has “transcription factor” as their secondary function belong to this set. The second most prevalent pair is enzyme/adhesion protein for pathogens (48 out of 288).

MOONPROT (http://moonlightingproteins.org/) is a database compiled by Jeffery et al. that stores information about moonlighting proteins for which there exists biochemical or biophysical evidence [36]. It contains 291 proteins. MoonDB (http://tagc.univ-mrs.fr/MoonDB/) contains human moonlighting proteins recovered from the literature and candidates predicted by a protein-protein network-based approach (Chapple et al., in submission). These databases provide platforms for systematic analysis of multifunctional/moonlighting proteins.

2.7 Gene Ontology Annotations of Moonlighting Proteins

Most moonlighting proteins are found serendipitously by experiments. Consequently, the majority of these proteins are best known for their primary function. On the other hand, annotation in UniProt often lacks GO terms related to their moonlighting functions. Here, we show two such examples of experimentally known moonlighting proteins (Table 1). PFK1 (UniProt ID: Q92448) is an ATP-dependent phosphofructokinase that phosphorylates D-fructose 6-phosphate in the first committed step of the glycolysis pathway. Additionally, this protein has been found to be involved in rapid and selective degradation of peroxisomes by microautophagy [37]. In the PFK1 knockout mutant, peroxisomes are observed to remain outside of vacuole and degradation is disabled. The existing GO annotations for this protein includes 14 GO terms describing its ATP-dependent catalytic activity in glycolysis (5 in BP, 7 in MF and 2 in CC), but lacks GO terms describing the moonlighting function, “autophagy peroxisomes”. The second example is glutamate racemase (UniProt ID: D3FPC2). It is an essential enzyme in the cell-wall biosynthesis pathway in bacteria, because it converts D-glutamate to L-glutamate, an important building block for peptidoglycan synthesis. Independent of the enzymatic function, this protein in Mycobacterium tuberculosis is shown to have a role as an inhibitor of DNA gyrase [38]. The UniProt entry of this protein has eight GO terms that clearly describe its racemerase activity in cell wall biosynthesis (4 in BP and 4 in MF), but no GO terms regarding the moonlighting function (“DNA gyrase inhibitor”). In Table 1, we listed the GO terms for these two proteins from UniProt as well as GO terms we have chosen that describe the secondary function. As illustrated in these two examples, it is not rare that moonlighting proteins are well-annotated in terms of their primary functions but under-annotated regarding moonlighting functions.

Table 1.

GO annotations of PFK1 and murL

Protein Name/Uniprot ID a)	GO terms for Primary Function b)	Added GO terms for Moonlighting Function c)
PFK1/Q92448	GO:0006002-fructose6-phosphate metabolic process GO:0006096-glycolytic process GO:0008152-metabolic process GO:0016310- phosphorylation GO:0046835-carbohydrate phosphorylation GO:0000166-nucleotide binding GO:0003824- catalytic activity GO:0003872-6-phosphofructokinase activity GO:0005524- ATP binding GO:0016301-kinase activity GO:0016740-transferase activity GO:0046872-metal ion binding GO:0005945-6-phosphofructokinase complex GO:0005737- cytoplasm	GO:0016237- microautophagy GO:0010508- positive regulation of autophagy GO:0030242- peroxisome degradation
murL/D3FPC2	GO:0006807-nitrogen compound metabolic process GO:0008152- metabolic process GO:0008360-regulation of cell shape GO:0009252-peptidoglycan biosynthetic process GO:0008881- glutamate racemase activity GO:0016853- isomerase activity GO:0016855-racemase and epimerase activity, acting on amino acids and derivatives GO:0036361-recemerase activity, acting on amino acids and derivatives	GO:0008657- DNA topoisomerase (ATP-hydrolyzing) inhibitor activity GO:2000372- negative regulation of DNA topoisomerase (ATP-hydrolyzing) activity GO:0004857- enzyme inhibitor activity GO:0090143- nucleoid organization

^a)The name of the moonlighting protein and UniProt ID.

^b)GO terms in UniProt describing their primary function.

^c)GO terms we have added for their moonlighting function. No GO terms were found for the moonlighting functions in UniProt.

3 DISCUSSION

We have reviewed existing computational works on moonlighting proteins. These papers analyzed moonlighting proteins from several different perspectives, i.e. sequence-based function prediction, protein-protein interaction, and structural properties.

Generally speaking, one advantage of computational analysis is that it can provide a big picture of biological phenomena. However, because the number of known moonlighting proteins is still small, the existing works were based on datasets of limited size. Moreover, as we have pointed out, annotations in the database often do not reflect the moonlighting functions of these proteins. To enable large scale computational characterization of moonlighting proteins, a comprehensive online repository with consistent functional annotations is the foremost requirement. In this regard, the three databases which are currently available and under continuous development are a good resource for future studies.

Structural analysis can provide a physical concrete picture of moonlighting proteins. Although a drawback of structural analysis is that it is only applicable to proteins that have experimentally solved tertiary structures, it is noteworthy that computationally modeled structures could be used as structure prediction methods have matured in recent years [39–41]. To aid in finding binding sites of moonlighting proteins, methods for detecting binding pocket sites in protein structures [42] and predicting binding ligands [43,44] can be useful.

The mechanisms by which moonlighting proteins exhibit multiple functions differ from case to case. Ultimately, an integrative approach will be needed for comprehensive understanding and classification of moonlighting proteins, which combines various types of data, such as proteomics, phenotypes, genomics, and other biochemical data. Investigation of moonlighting proteins is still in its early stage. We foresee that moonlighting proteins will be more systematically studied in near future and anticipate that computational work will play important roles there.