LAB-Secretome: a genome-scale comparative analysis of the predicted extracellular and surface-associated proteins of Lactic Acid Bacteria

Construction of the secretome protein clusters (LaCOGs)

In this study we focus on those proteins that are predicted to be wholly or largely on the outside of the cell, regardless of the translocation systems. These proteins form a sub-proteome of what Desvaux et.al defined as the "secretome" [27] by excluding the translocation systems, the integral membrane proteins, and non-protein products. Although we adapt this term "secretome" to describe our protein subset of interest, we must specify that in our analysis the term "secretome" refers to only the proteins that are released from the cells to the extracellular milieu (also called exoproteome), and the proteins that remain cell-surface associated, but nothing else.

Ideally, a comparative secretome analysis should be performed on the experimentally validated sub-proteomes or on in silico predicted secretome proteins with the highest possible accuracies. However, it is well-known that wet-lab proteomic studies are extremely costly and can lead to many false predictions of subcellular location, while all the currently available in silico protein SCL predictors have only 80%-93% prediction accuracy [38–41]. Therefore, instead of clustering predicted extracellular proteins directly, we designed an alternative process which firstly groups all proteins in the sequenced LAB genomes into ortholog groups (LaCOGs) and afterwards extracts the secretome groups by using genome-scale SCL predictions (Figure 1). In this way, the wrongly predicted secretome proteins could be reduced because homologous proteins with similar functions and domains always tend to have the same SCL, and vice versa[39–42].

The Lactobacillales-specific clusters of orthologous groups of proteins (LaCOGs) previously generated by Makarova et.al[43] were used as the basis for protein clustering into protein families. In total 3374 (729 new and 2645 existing) LaCOGs were formed by adding 14 recently sequenced LAB genomes to the Makarova et. al. set. Subsequently, a genome-scale SCL prediction was performed on all proteins in the 26 genomes (Table 1). By combining the SCL prediction and LaCOGs, and after manual curation (see below), we defined 462 secretome LaCOGs (of which 212 are new compared to the Makarova et. al. set) composed of 3357 proteins, representing 7.4% of the complete genome dataset and 93% of all predicted secretome proteins in these 26 genomes. We defined thirteen general functional classes for these proteins, and the distribution of these clustered secretome proteins over the classes and LaCOGs is shown in Figure 2. An additional 249 putative secretome proteins could not be grouped into these LaCOGs, comprising 69 proteins that had only a distant homolog in non-LAB, and 180 proteins that had no homolog in any sequenced bacterial genomes, which we termed the extracellular "ORFans" (Table 2, Additional file 1, sheet S1).

LAB species and strains	Secretome size	Proteins in LaCOG	Distant Homologs	ORFans	LaCOGs
E.faecalis V583	281	232	22	27	131
L.acidophilus NCFM	171	161	2	8	108
L.brevis ATCC 367	177	154	5	18	113
L.casei ATCC 334 *	192	187	3	2	148
L.casei BL23	205	197	0	8	153
L.citreum KM20	112	112	0	0	93
L.delbrueckii bulgaricus ATCC BAA-365 *	115	113	0	2	94
L.delbrueckii bulgaricus ATCC11842	87	79	3	5	68
L.fermentum IFO 3956	112	112	0	0	89
L.gasseri ATCC 33323 *	115	113	0	2	88
L.helveticus DPC 4571	131	123	2	6	97
L.johnsonii NCC 533 *	236	209	6	21	131
L.lactis cremoris MG1363	105	103	0	2	86
L.lactis cremoris SK11 *	105	105	0	0	87
L.lactis lactis IL1403 *	136	114	4	18	80
L.mesenteroides ATCC 8293 *	112	94	5	13	77
L.plantarum WCFS1 *	160	151	5	4	123
L.reuteri F275 JGI	159	156	1	2	124
L.reuteri F275 Kitasato	171	156	2	13	123
L.sakei 23K	114	103	4	7	80
L.salivarius UCC118	135	126	3	6	103
O.oeni PSU-1 *	95	90	0	5	70
P.pentosaceus ATCC 25745 *	99	89	1	9	79
S.thermophilus CNRZ1066 *	90	90	0	0	77
S.thermophilus LMD-9 *	97	94	1	2	84
S.thermophilus LMG 18311	94	94	0	0	81

Although the LAB genomes vary in size, the size of the secretome as a fraction of each genome was fairly consistent (6-10%), as well as the distribution of proteins over different SCLs. The N-terminally anchored proteins with no signal peptidase cleavage site are the most abundant kind among all predicted secretome proteins. A striking feature of numerous secretome proteins, and particularly surface-associated proteins, is that they are large and consist of many different domains (often in repeats), and domain compositions (see examples in Figure 3). In fact, this variation in domain composition has been used in constructing and sub-dividing the LaCOGs and separating sub-families of homologous proteins. Distinct combinations of domains provide hints for functions of these extracellular proteins in cell-wall metabolism, cell-wall binding and their communication with the environment (see below).

False predictions and pseudogenes

The preliminary secretome clusters were curated manually and corrected based on expert knowledge, e.g. for false-positive and false-negative predictions, incorrect gene starts, pseudogenes, etc. Examples of proteins of known intracellular function, but with consistent false-positive extracellular SCL prediction are listed in Additional file 2, sheet S1. In most cases the mis-prediction was caused by an α-helix-like N-terminal sequence in these proteins (possibly as part of the hydrophobic core of a globular protein), leading to the prediction as a signal peptide by LocateP. A further improvement was made by finding and removing those LaCOGs that have proteins which are anchored in the cell membrane with a single N-terminal transmembrane helix, but with the rest of the protein inside the cell (so-called outside-in topology, GO:0031235) [44–53]. By aligning proteins within these LaCOGs we found that these proteins do not have positively charged residues preceding the N-terminal hydrophobic helix, but exclusively have a positively charged residue(s) immediately downstream of the transmembrane helix (examples in Additional file 2, sheet S2). Hence such features could be used for further development of a model for SCL prediction of N-terminally anchored proteins by LocateP.

Nearly 400 pseudogenes were identified, but this is probably an underestimate. In most cases this was due to gene frameshifts, and occasionally to N- or C-terminal truncation of genes. Most of these genes could be concatenated to encode larger proteins with high similarity to known proteins in the LaCOGs. Many of these pseudogenes were initially predicted to encode intracellular proteins by LocateP, but after concatenation these proteins are predicted to be extracellular and/or contain domains of extracellular functionalities. An example are the proteins encoded by adjacent genes LSA1731 and LSA1730 in L.sakei 23K which were annotated as hypothetical proteins. The concatenated protein showed high similarity to proteins in LaCOG02935 which were exclusively cell-surface protein Csc complex family members [54]. In total 129 concatenated pseudoproteins were made with 279 protein fragments (Additional file 3, sheet S1), while 87 pseudogenes could not be combined (Additional file 3, sheet S2).

The LAB-Secretome database

The LAB-Secretome database http://www.cmbi.ru.nl/lab_secretome was constructed to store and browse all the predicted extracellular proteins and LaCOGs. An overview page summarizes all predicted secretomes, LaCOGs, distant homologs in non-LAB species and the ORFans, with hyperlinks to the corresponding HTML pages to help users to browse the whole database (Figure 4A). The LAB-Secretome database can be queried in many ways, e.g. by bacterial species, protein subcellular location, protein accession identifiers, LaCOG numbers, protein functional classes, and Pfam domain accession codes or domain functions (Figure 4B). Visualization includes a description of LaCOG members and function, protein functional domain composition, and multiple alignments with notification of corrected start codons, pseudogenes and concatenated proteins (Figure 4D). A Blast function, utilizing the BlastP [55] program, enables users to query the clustering information of their proteins of interest to the extracellular proteins and families that are already in the database (Figure 4C). An automatic updating scheme for the LaCOGs (Figure 1) was designed to ensure that the need for manual curation is minimized when adding new bacterial genomes to the database.

Overview of the extracellular protein families

Most common functionalities in the secretomes of LAB

Among all 215 secretome LaCOGs with known or presumed functions, almost half of them contain proteins which are involved in cell-wall metabolism, e.g. the muramidase, lysin, lysozyme and beta-lactamase families (Figure 2). Many of these enzyme families are further subdivided into different LaCOGs based on variations in sequence homology and protein domain compositions, and some may represent species/niche-specific subfamilies. One example is the subdivision of proteins with an Nlpc/P60 family domain (e.g. gamma-D-glutamate-meso-diaminopimelate muropeptidase) into 5 separate LaCOGs (Additional file 4, sheet S1). These proteins vary in length from ~150 to ~500 amino acids, all with the Nlpc/P60 domain in the C-terminal part. In only one of these subfamilies (LaCOG90015), all 16 members have 1-3 copies of LysM domains (Pfam PF01476) in their N-terminal part, indicating extra binding functions to the cell-envelope. A similar domain architecture is found in one of the four N-acetylmuramoyl-L-alanine amidase subfamilies (LaCOG01848), which has an enzymatic C-terminal domain and 0-3 N-terminal SH3 domains (Pfam PF08239), known to bind to proline-rich regions of proteins. In the pepdidoglycan hydrolase subfamilies LaCOG00186 and LaCOG01653 the enzymatic domain is located at the N-terminus and can be followed by different kinds, combinations and numbers of binding domains such as LysM, SH3 or surface layer domain (Pfam PF03217) (Figure 5). These examples all illustrate that the many types of extracellular enzymes involved in cell-wall turnover have different mechanisms to attach to components of the cell surface.

Specific enzyme families

Specific binding-protein families

Many extracellular proteins contain known domains for binding to macromolecular substrates. In addition to domains for binding to the cell wall of the producing cell (e.g. LysM, SH3), several other domains are found which are related to binding to host macromolecules (e.g. domains annotated as mucus-binding, chitin-binding, collagen-binding, fibronectin-binding, carbohydrate-binding, etc) (Table 3). Some of these annotations derive from in vitro binding studies and may not reflect in vivo functions. In LAB, mucus-binding domains (MUB, MucBP) are found in many proteins and are thought to play a role in binding to the host GI-tract mucus layer [57, 93, 94]. An enormous variety is found in the size of these mucus-binding proteins and in the number of mucus-binding domains. We have made a preliminary separation into 7 different subfamilies of mucus-binding proteins based on protein size, sequence homology, domain composition and phylogeny (Table 3). The three largest subfamilies are (1) LaCOG00885 containing 11 members from different LAB but not from L. acidophilus group members, (2) LaCOG01470 with 28 members, found in many LAB, and (3) LaCOG03211 which includes 10 proteins. The proteins of LaCOG00885 contain solely the MucBP domains as defined by Pfam (PF00746), while the proteins of the other two LaCOGs possess multiple copies of the larger MUB domains as defined by Boekhorst et al.[71] (see also Figure 2 in[95]). Many mucus-binding proteins of L. acidophilus group members contain an N-terminal [Y/F]SIRKxxxGxxS-containing signal peptide (PF04650) which was earlier reported as a typical characteristic of the L. acidophilus MUB proteins [94, 96], and may relate to a specific function in sorting or folding [97, 98]. Furthermore, it is striking that many large genes encoding mucus-binding proteins are pseudogenes (e.g. in LaCOG01470, LaCOG03211 and LaCOG99309). While it is unlikely that these are all due to sequencing errors, it is not clear yet whether these are truly pseudogenes, or possibly may encode functional proteins after transcription with strand-slipping [5, 71].

Genome sequences and bioinformatics tools used in this research

The genome sequences of 26 selected representative lactic acid bacteria, including the protein functional annotation and the gene contexts, were obtained from the NCBI bacterial genome database (version 15 Aug., 2008) [114].

BlastP (default cutoff values of E < 1, low-complexity filter disabled) [55] and Inparanoid [115] were used for sequence homology and orthology searches, respectively. Protein subcellular location (SCL) was predicted by LocateP [38]. Multiple sequence alignments were constructed using Muscle [116]. Motif searches were performed using MEME and MAST [117]. Protein domains (version Dec. 2008) [118] originating from the Pfam database [119–121] and additional HMMs reported in other studies [54, 71, 96, 122–124] were searched using HMMER [125] with the respective cut-off of each model. The domain functions were obtained from the GO database [126] using the PFAM2GO dataset [126].

The LAB-Secretome database was created in MySQL and the database interface was written in PHP (version5.2.7). Visualization of the protein domain composition was made using scalable vector graphics (SVG).

Protein clustering into orthology groups (LaCOGs)

First, the 22,191 proteins in 3195 LaCOGs generated by Makarova et.al[1] from 12 LAB genomes were used as the basis for protein clustering. All protein sequences from 14 newly sequenced LAB genomes were searched against the Makarova LaCOG set using BlastP. The proteins that have high homology to the existing LaCOGs were then selected using a revised criterion based on the well-known COG extension rule "uniform top 3"[127]: if all the top 3 (in case of LaCOG size of 2, the top 2 hits were taken) BlastP hits of a query protein belong to the same LaCOG (LaCOG size bigger than or equals to 2), then the query protein is added to this LaCOG.

Since the above-mentioned extension was purely based on the homologs of proteins that were already included in the LaCOGs by Makarova et al., the specific proteins from newly sequenced species, e.g. L. reuteri, were not added due to the absence of the "seeding sequences" for BlastP. In order to cluster all proteins that originated from the newly sequenced genomes, a complete all-to-all Inparanoid [115] search was performed in a parallel fashion with the proteins encoded in the 14 new genomes to identify orthologous proteins. Cut-off settings of bit score 50 and sequence overlap of 50% were used. The proteins with all-to-all bidirectional-best-hit (BBH) relationship [128, 129] were clustered into groups, meaning that in any such group, each member is the BBH of another member. This stringent criterion generates new cores of orthologous proteins.

Using the core ortholog clusters and the extended LaCOGs made in step one, the proteins that were not previously included in any clusters, including those proteins from Makarova LaCOGs containing only 1 member, were Blasted as queries. In this step, the revised criterion "uniform top 3" was used and new LaCOGs were made.

The newly made LaCOGs were merged with the extended Makarova LaCOGs, and the newly made ones were assigned coding numbers starting with "9" in their names, e.g. LaCOG90001, to distinguish them from the extended Makarova LaCOGs.

LaCOG quality control

In order to check the quality of the merged LaCOGs, an iterative BlastP search was performed using the clustered proteins as queries against all the proteins that were not included in any constructed LaCOGs, using the criteria of 1E-3 and query-hit protein length ratio of 0.6, which has been tested by Boekhorst et. al.[130] for distant homolog identification. This iterative search found that only 13 non-clustered proteins (mostly hypothetical proteins) had a distant homolog in 11 different LaCOGs, indicating that our clustering methods have extensively included most of the proteins into possible homologous clusters.

ORFans and proteins with only non-LAB distant homologs

The LAB proteins that could not be clustered into LaCOGs by the previously described procedures were then collected and Blasted against all completely sequences non-LAB bacterial genomes (both Gram- and Gram+ species). The same criterion of distant homolog identification [130] was utilized. Proteins that had no homologs in any other species were named "ORFans".

Secretome LaCOG extraction

The clustering information of merged LaCOGs, proteins that have only distant homologs in non-LAB species and the ORFans was then combined with the SCL prediction made by LocateP (Table 1). Initially, only the LaCOGs that had at least half of the members with a predicted secretome SCL corresponding to (1) lipid-anchored; (2) N-/C-terminally anchored; (3) secreted by Tat- or Sec- pathway; (4) secreted via non-classical pathways, or (5) cell-wall anchored were identified as the secretome LaCOGs. Later, all other LaCOGs were manually inspected, and a few families were identified with a mixture of secretome and intracellular proteins; only the secretome proteins were added to the database. The same classification was applied to the secretome ORFans and proteins that have only distant homologs in non-LAB species. The resulting clusters of secretome proteins, the "secretome", can be further extended by similar processes when new (LAB) genome sequences become available.

Proteins that are exported by unknown mechanisms and so-called "moon-lighting" proteins (known intracellular function, but often also found on the outside of the cell) [131] were not considered as their extracellular SCL cannot be predicted.

Manual curation

In order to obtain as accurate as possible prediction of secretome proteins and their classification into LaCOGs, we performed a throughout manual inspection on all the secretome proteins, including the ORFans and the ones included in LaCOGs. All proteins were double checked for the ORF-calling quality by the criteria combining protein length, possible alternative start (end) codon, multiple sequence alignments, protein domain composition and SCL prediction consistency.

Incorrectly chosen start codons in the original annotations were corrected based on sequence alignment with protein family members, position of putative ribosome-binding sites, and known features of signal peptides. Pseudogenes were initially identified when BLASTP analysis of the encoded proteins showed that they belong to extracellular protein families in LaCOGs, but that they represented only a fragment of the protein. By analysis of the coding region of these pseudogenes with their adjacent nucleotide sequences we could generally identify frameshifts, such that the missing protein part(s) were found to be encoded in a different reading frame. In these cases, the entire opening-reading frames were translated into protein fragments, regardless of the absence of start codons, and these protein fragments were concatenated to form new protein sequences that share high similarity to other known full-length proteins. In a few cases, ORFans were also identified as pseudogenes when they lacked a signal peptide, but otherwise contained protein domains typical of extracellular proteins.

Generally, we expected the ORFans to be real genes that represent unique functionality to the specific LAB in which they occur. However, because the average size of these hypothetical ORFs was below 100 amino acids, it is possible that some small ORFans could as well be wrongly predicted ORFs or pseudogenes. Proteins smaller than 80 amino acids containing only a Sec-type N-terminal signal sequence were removed from the set of predicted extracellular proteins, since their C-terminal part is generally too small to represent an extracellular domain. Moreover, many of such small proteins with a single predicted TM helix are now increasingly considered as small integral membrane proteins [132].