Tools for Exploring the Proteomosphere

Abstract

Homology-driven proteomics aims at exploring the proteomes of organisms with unsequenced genomes that, despite rapid genomic sequencing progress, still represent the overwhelming majority of species in the biosphere. Methodologies have been developed to enable automated LC-MS/MS identifications of unknown proteins, which rely on the sequence similarity between the fragmented peptides and reference database sequences from phylogenetically related species. However, because full sequences of matched proteins are not available and matching specificity is reduced, estimating protein abundances should become the obligatory element of homology-driven proteomics pipelines to circumvent the interpretation bias towards proteins from evolutionary conserved families.

Introduction

Proteomics has become an integral part of molecular and cell biology (reviewed in [1-3]). One, however, should admit that, despite remarkable progress in both mass spectrometry technologies and bioinformatics, the significance of proteomics methodologies for biological research has been granted by the pace, precision and scope of sequencing and interpretation of genomes.

In a typical LC-MS/MS or MALDI MS/MS experiment thousands of tandem mass spectra are acquired from the majority of detectable peptide precursor ions. Computational post processing converts MS/MS spectra into peak lists, which are then submitted to database searches by the dedicated software (reviewed in [4-7]). Irrespectively of its algorithmic base, the software fulfils two major functions: first, it provides a numerical estimate of the likeness of a particular MS/MS spectrum to its in silico representation computed from the database peptide sequence using empirical fragmentation models. Second, it evaluates their similarity in respect to random matches and, hence, computes the estimate of probability that this match is correct. Despite numerous algorithmic, statistical and bioinformatic limitations, this is an efficient and generic approach that underpins the majority of today's proteomics efforts.

We underscore that a typical database search does not really interpret the spectrum. Instead, it tentatively annotates the observed fragment ions and ultimately relies upon the scored similarity between the observed and predicted patterns. Not surprisingly, a very high specificity of spectrum-to-sequence correlation requires obeying stringent matching constraints. Any discrepancy between the actual and database sequences – regardless, if it originates from an amino acid substitution, unexpected post-translational modification or peculiar fragmentation pathway, might deny the peptide identification completely. To produce a hit, the database search would then require more permissive (and, consequently, much less specific) error-tolerant searches. Therefore, conventional proteomics approaches favor the organisms for whom large and accurate database sequence resources or fully sequenced genome are available.

These organisms, however, are not even remotely representing the full biosphere complexity, while much can be learned by exploring a proteomosphere in its entirety. Indeed, classic biochemical and enzymological techniques help to purify and assay interesting activities regardless if the exact sequences of corresponding protein factors are known. The prospective lines of research include (yet, by far, are not limited to) the adaptation mechanisms in animals and plants; infection of plants by viruses and pests; exploring venoms and natural biofluids as a source of pharmacologically important enzymes and proteinous bioregulators, among others.

Challenges and expected benefits of expanding the organismal scope of proteomics have been debated [8-10]. Rapid increase in size and organismal representation of sequence databases suggested that a sizable number of proteins from any free living species should be identifiable on the basis of their similarity to already known protein sequences. Although much has been accomplished in both technology developments and applications, new challenges have emerged. In this review, we would like to re-access the achievements and challenges of the homology- driven proteomics and discuss possible instrumentation and bioinformatic solutions.

Technical aspects of homology – driven protein identifications

A typical scenario in homology-driven proteomics considers the identification of unknown (i.e. not present in a database) proteins, assuming that a database resource contains distantly homologous protein sequences from phylogenetically related species. Therefore, the analysis more commonly hits a family of homologous proteins in several species, rather than a unique protein.

Homologous proteins are identified by tandem mass spectrometry via several data mining strategies (reviewed in [11]) (Figure 1).

Figure 1.

Matching peptide sequences in homology-driven proteomics. The choice of optimal identification algorithm depends on the similarity between analyzed and reference proteins, as well as the quality of available MS/MS spectra. X symbol stands for any amino acid residue. The local dissimilarity regions are boxed.

If the analyzed protein belongs to a conserved protein family, it is likely to comprise several peptides that are fully identical to reference sequences in a database. Its cross-species identification does not differ from the identification of known (i.e. present in a database) proteins by conventional (stringent) searching means, albeit the achieved sequence coverage might be compromised. The more peptides are sequenced in MS/MS experiments, the larger dissimilarity to the reference proteins can be tolerated, while protein assignments would be gradually deviating from the specific protein hit to a larger protein class sharing only a few identical peptide sequences.

If specifically requested, several conventional algorithms can also tolerate a single amino acid mismatch within the compared peptide sequences or disregard the enzyme cleavage specificity (Figure 1). Protein identification still proceeds with uninterpreted MS/MS spectra. However, because of alleviated matching specificity constraints, the search space dramatically increases and lengthens database searches. Moreover, maintaining the same false discovery rate requires to employ higher confidence thresholds, which concomitantly increases the false negative rate.

The concept of sequence tag searches (Figure 1) assumes that a single mismatched piece of peptide sequence (regardless of its length) is adjacent to the N- or C-terminus of the analyzed peptide, while the rest of its sequence is identical to the peptide in a database. A peptide sequence tag – a sequence stretch of, typically, 2 to 4 amino acid residues - could be deduced from the MS/MS spectrum and, together with m/z of the corresponding fragment ions, employed in error-tolerant searches [12]. One part of the tag – either the sequence itself or one of the two flanking m/z values -is allowed to mismatch. Therefore, error-tolerant searches typically produce large hit lists requiring mandatory manual inspection. The search specificity could be improved and manual inspection alleviated, if hit lists of multiple error-tolerant searches with several sequence tags produced from each of the fragmented peptides are combined and evaluated together using a dedicated statistical model [13]. It is also possible to automatically deduce tags and submit them to database searches [14, 15]. Compared to a single amino acid residue mismatch, sequence tag searches are far more permissive, although they do require that matched peptides share a sizable piece of identical sequence. Importantly, it is usually straightforward to retrieve short sequence stretches even from marginal quality MS/MS spectra, which are not amenable to accurate de novo interpretation. This improves the sensitivity and sequence coverage of the analysis because it encompasses more spectra of diverse quality. However, sequence tag searches are not much used in homology-driven proteomics, probably because no integrated software is available that could cover all aspects of the protein identification pipeline starting from generating tags till estimating the confidence of database searching hits despite algorithmic solutions have been developed for each of its individual elements.

Alignments of full length sequences of analyzed peptides with reference sequences could tolerate multiple mismatches and sequence stretches identical between the two compared peptides can be short and scattered within the peptide backbone [16] (Figure 1). Yet, it is not straightforward to obtain sequences amenable for such comparison. Despite high mass accuracy and mass resolution of instruments [17] that advance spectra interpretation algorithms [18-20], de novo sequencing still represents a considerable challenge. Tandem mass spectra are reflecting sequences of fragmented peptides, spectra are not representing sequences accurately and completely, while certain m/z regions are devoid of useful fragments. It is often possible to retrieve a few pieces of peptide sequence out of any MS/MS spectrum (reviewed in [10, 21]), yet they are usually ambiguous and composed of intermingling correct and false stretches. Therefore, de novo interpretation typically yields a number of sequence candidates (rather than a single unique sequence), which are usually short (10-15 amino acid residues), redundant, degenerate and partially incorrect [22, 23]. Their relative location at the backbone of analyzed protein is not known and, furthermore, peptides sequenced in the same experiment might originate from different proteins. Although algorithms for aligning protein sequences and registering their homology have been perfected by bionformatics over decades [24, 25], peptide sequence candidate produced by automated de novo interpretation of MS/MS spectra can not be fed to BLAST or FASTA engines directly without compromising search specificity. Several database searching tools, which handle large numbers of de novo sequence candidates have been developed [26-29]. Some of them are web-accessible, like (MS BLAST) [16, 30] or FASTS [29], or are now available as a part of commercial software packages (ProBLAST from Applied Biosystems [31]; MS BLAST option in BioTools from Bruker Daltonics).

Sequence similarity searches employ peptide sequence candidates (rather than raw MS/MS spectra) and are not influenced directly by the instrument type. It is only important that sequence predictions are reasonably accurate (although several sequence variants per each interpreted spectrum are allowed) and de novo sequencing output is formatted into a query string according to a few simple conventions. Large search queries are allowed: while they could contain candidate sequences deduced from an unlimited number of interpreted spectra, their full sizes are limited by 150,000 amino acid resides in total, which is roughly equivalent to a BLAST search with 16.5 MDa protein chimera [30]. Therefore, it has become technically possible to combine automated sequence similarity searches with sensitive LC-MS/MS analysis, which used to produce thousands of MS/MS spectra per each run. Now it might be also practical to subject all acquired spectra to de novo interpretation: besides sequence similarity searches, de novo sequencing also assists in validating “conventional” peptide hits having marginal statistical confidence [32, 33].

However, we see several reasons why “full de novo sequencing” approach, if taken carelessly, might be misleading.

One of the major concerns is that poor quality MS/MS spectra might produce fully incorrect interpretations and, to control the false positive rate, such spectra should be removed prior to database search by some quality filter [34, 35]. This, however, appears to be less of a danger. Any de novo sequencing software offers an arbitrary quality score, which reflects the confidence of produced sequence candidates. Using computational experiments with simulated spectra datasets, it is relatively straightforward to determine arbitrary cut-off scores, below which the produced sequences have almost no chance of hitting correct target proteins in error-tolerant alignments (although even very high de novo quality scores do not guarantee that this peptide(s) will produce a hit). Hence, unreliable sequence candidates could be sorted out based on their arbitrary de novo confidence and, if necessary, added back to the search query later to increase the sequence coverage of already made confident hits or to consider a few more marginal alignments.

Second, by far more serious, concern relates to the reduced matching specificity of sequence similarity searches. A large fraction of MS/MS spectra acquired in a typical LC-MS/MS run under automated data-dependent control could be directly attributed to peptides originating from background proteins – trypsin, human keratins, cell medium components, etc and even more of them could be matched via error-tolerant searches [36, 37]. While peptides from background proteins do not harm conventional protein identifications, they strongly affect sequence similarity searches by hitting a very large number of totally unrelated proteins to which they might bear some local similarity. For example, trypsin autolysis products used to hit a multitude of serine proteases of diverse functional specificity and species of origin. Because of low complexity sequence stretches, keratin peptides could match almost any protein in a database with varying degree of statistical confidence. We note that conventional database searches, performed prior to de novo interpretation of spectra only remove a relatively small fraction of background spectra [36] with little effect on sequence similarity hit lists.

As a practical solution, it was proposed to subtract such MS/MS spectra by comparing them to a library of non-annotated background spectra, compiled from control and blank LC-MS/MS runs [36, 38]. Spectra screening software is available as a stand-alone application, or at the public web server [30]. Depending on the relative abundances of background and target precursors, filtering removes up to a half of the total number of acquired MS/MS spectra, while no high quality spectra originating from bona fide target proteins are lost. Importantly, filtering does not rely on sequence database resources and exactly the same algorithm equally applies for processing spectral datasets acquired from known or unknown proteins [36].

We underscore that sequence similarity searches tolerate the compromised accuracy of candidate sequences and therefore no chemical modification of peptides that support de novo interpretations [39-44] is required in most cases. Therefore, the same data acquisition and sample processing routines could be employed at the uncompromised detection sensitivity, regardless if the analyzed sample might contain known or unknown proteins. For most efficient analysis, multiple datamining routines could be combined into a “layered” proteomics pipeline, which supports flexible manipulation with individual MS/MS spectra [28, 30].

Taken together, it is reasonable to assume that major technical issues in homology-driven protein identifications have now been addressed. Convenient workable software solutions for data pre-processing, de novo sequencing and sequence-similarity searches are available and could be integrated into automated proteomics pipelines supporting any instrument type with adequate MS/MS capabilities.

Scope and applications of homology driven proteomics

The success rate of cross-species identifications depends on three major factors: first, on the phylogenetic distance between the analyzed and the reference organism(s); on the proteome coverage of the reference organism in a database and on the number of peptides sequenced from the analyzed protein. Computational modeling suggested that, using a dataset of 10 to 15 de novo sequenced peptides, sequence similarity searches should be able to match 50 to 60% of proteins having at least 50% of the sequence identity with the reference proteins [45]. Furthermore, because of relatively small phylogenetic divergence between individual species and availability of several completed genomes, it should be possible to identify > 75% of proteins within mammalian and plant kingdoms on the proteome –wide scale [45, 46]. LC-MS/MS sequencing of 10 to 20 precursors is usually unproblematic if the target protein is present at the low picomole level and therefore the above estimates of proteome coverage scope are certainly within the practical reach.

We note, however, that in phylogeneticaly distant species the similarity of protein sequences does not necessarily imply their functional similarity. Therefore, using homology- driven protein identifications for ontology annotations of corresponding protein products should be performed with great caution [47, 48].

There are numerous examples of successful application of sequence-similarity identifications in plant and animal biology, microbiology, toxicology, environmental studies, to mention just a few areas (reviewed in [49-51]). Whereas most abundant proteins typically come from conserved protein families and are readily identified by conventional searches, in our experience sequence similarity searches usually add more than 25% of new identifications [46, 52-58]. By matching more peptides, sequence-similarity searches increase the sequence coverage and identification confidence of hits made by conventional searches, which is helpful in distinguishing protein isoforms and/or individual members of protein families.

Among wild-bred species pronounced polymorphism of protein sequences occurs [59]. Even in a digest of the same protein spot, it is not uncommon to detect several sequence variants of the same tryptic peptide, which differ by a single amino acid between themselves and between the database reference sequence (see, for example [55]). Although polymorphism effectively precludes their identification by conventional database searching means, registering sequence commonalities by sequence-similarity searches is usually unproblematic.

In conclusion, we note that the progress in EST and genomic sequencing enhances the scope and practical usability of homology-driven proteomics, rather than making it redundant and obsolete. First, since novel reference genomes appear from previously underrepresented phylogenetic kingdoms, more and more species and classes become amenable to proteomic characterization [9]. Second, homology-driven proteomics effectively bridges reference genome sequences with the natural sequence variability of wild living species.

Protein Quantification: Obligatory Element of a Homology-driven Proteomic Pipeline

Identification confidence is not directly affected by the available protein quantity and several database search engines disregard the actual abundance of precursor and/or fragment ions, solely operating with normalized peak intensities. While analyzing protein mixtures, we also presume that expert consideration of physico-chemical properties of proteins and fine tuning data acquisition and database searching settings will result in adequate protein representation in the hit lists. If necessary, relative quantities of individual protein components could be estimated using the abundances of matched peptide precursors or by MS/MS spectral counts or related indices (reviewed in [1, 60-62]).

This, however, is only true if complete sequences of analyzed proteins are present in a database. In a mixture of known and unknown proteins conventional searches will only identify the former, despite the latter protein might constitute the major component. In the case of stringent searches, higher abundance of precursor peaks typically translates into a better quality of corresponding MS/MS spectra and they are more often matched, rather than falsely mismatched. However, if the analysis identifies a protein with unknown sequence by a few matched peptides, it is quite common that several high quality spectra still remain unmatched. It is not possible to tell if they originate from an uncovered piece of sequence of the already identified protein, or they indicate that major protein component(s) remain unidentified by all database searching means.

The case study presented below exemplifies the problem severity. The abundance of a silver-stained spot with apparent MW of 38.1 kDa and pI 5.42 visualized on a two-dimensional gel of the protein extract from leaves of the European beech Fagus sylvatica was decreased upon infecting the tree with root pathogen Phytophthora citricola. The corresponding spot isolated from control and infected plants were excised from the gels, digested with trypsin and analyzed by LC-MS/MS on a hybrid LTQ Orbitrap mass spectrometer. MASCOT searches produced four confident cross-species hits (Table 1), while homology-driven proteomics pipeline [30] additionally identified an RNA binding protein based on its homology to known Arabidopsis thaliana and Oryza sativa proteins (Figure 2).

Table 1.

Proteins identified in a spot from Fagus sylvatica by MASCOT and MS BLAST searches

	Protein	Accession no.	MASCOT search *			MS BLAST search
			Total Score	Matched spectra	Unique peptides	Sequence coverage, %	Total Score	Matched spectra	Unique peptides
GTM	γ tocopherol methyltransferase	AAX63740	160	4	4	13
Pdx	Pyridoxine biosynthesis protein	NP_195761	110	2	2	8	82	4	1
LHCP	Light harvesting chlorophyll a/b binding protein	CAA48410	73	2	2	8
USP	Universal stress protein (USP)	BAA97516	68	1	1	4
RBP	Putative RNA binding protein	NP_181084					269	18	4

^*searches were performed against NCBI; 1 missed cleavage; Mass tolerance: precursor ions: 10 ppm; fragment ions: 0.6 Da; Fixed modifications: Carbamidomethyl (C); Variable modifications: Acetyl (N-terminal), Oxidation (M), Phosphorylation (S,T,Y).

Figure 2.

Identification of RNA binding protein in the same spots that were excised from control and treatment gels by LC-MS/MS, automated de novo sequencing and MS BLAST search. Score were assigned to corresponding high scoring segment pairs (HSPs) by the BLAST engine.

There is no robust methodology to determine the abundance of proteins in a mixture without internal standards [63, 64]. Therefore we further followed the idea of Silva et al [65] and, solely as a ballpark estimate, inferred the relative amounts of proteins from the normalized intensities of most abundant precursors matched to correspondent sequences (Figure 3). According to our estimate, RNA binding protein – only identified by MS BLAST searches - represented a major component of the spot, which was further supported by a large difference in MS/MS spectra count indices, which was 4.5 for peptides from RNA binding protein compared to 1 for peptides from other proteins identified in the spot digest. Its abundance dropped following the pathogen infection and is probably responsible for the apparent decrease in the spot intensity, despite the abundance of other minor components increased (Figure 3).

Figure 3.

Differences in the relative abundance of five protein components identified in the silver stained spot (designated with the pointer at the inset images) in protein extracts of control and infected plants. Full protein names are provided in Table 1.

Hence, even qualitative interpretation of the LC-MS/MS analysis ultimately requires to determine if it identified the major component(s) in the sample, or only minor, yet relatively conserved proteins were hit. In our experience, this is the most serious bottleneck of the current homology-driven proteomics efforts. Effectively, it requires quantifying proteins with unknown sequences and relating most abundant, yet unmatched, precursors to identified components. Although it might be impossible to offer a generic solution, we could think of several empirical approaches partially addressing these issues. For example, quantitative comparison of peptide profiles acquired in mixtures several independent experiments might reveal coherent changes in abundances of certain precursor peaks [66], irrespectively if they have (or have not) been matched to identified proteins. Alternatively, large species-specific libraries of non-annotated reference MS/MS spectra could be composed by the systematic analysis of samples with known content and used for identifying newly acquired spectra independently of sequence database resources [36, 38, 67, 68].

Conclusions

Homology-driven proteomics is a mature field that enjoys steady expansion of its application scope. By inferring identities of unknown proteins, it provides insight into the molecular mechanisms of complex biological phenomena that were previously addressed by descriptive phenomenological approaches. It has become possible to determine the representative molecular composition of proteomes of biological fluids, tissues or organs obtained from a multitude of free-living species and quantify their dynamic response to a variety of endogenous or environmental stimuli. The techniques supporting sequence similarity identifications are generic and can be incorporated into any proteomics pipeline. We anticipate that, in meantime, sequence similarity identifications should become routine and commonly used by proteomics laboratories studying organisms with unknown genomes and / or wild-bred species. The reviewing policies might request the obligatory use of both stringent and error-tolerant searches as a prerequisite of adequate representation of the reported proteome composition.

It has also become obvious that homology driven protein assignments should be supported with some quantitative evidence, since sequence-similarity interpretations are inherently biased towards matching known or most conserved proteins. On the positive side, massive efforts in developing label-free quantifications methods are on-going and corresponding software algorithms for unbiased extraction of precursor intensities will be of considerable value for validating homology-driven identifications.

Despite apparent significance of homology-driven proteomics for pharmacological, environmental, conservational, agricultural, among many other fields of biological science, there might be some less explored, yet intriguing opportunities. In particular, it seems exciting to build a tighter alliance with developmental and evolutionary biology. In our (may be, somewhat biased) opinion these field “stagnate” within the realm of established model organisms amenable to genetic manipulations. Yet, it is conceivable that studying the cellular molecular machinery on a wider organismal scope might dramatically improve our understanding of the evolution and adaptation mechanisms.