Special Issue of Mass Spectrometry Reviews on "Computers in Mass Spectrometry": “Proteogenomics from a bioinformatics angle: a growing field”

1. Introduction

Proteogenomics integrates two different research fields, mass spectrometry (MS) based proteomics and next-generation sequencing (NGS) based genomics, transcriptomics or translatomics. At the outset, proteomics data was used to aid genome annotation by gene model refinement based on protein-level validation¹. Therefore fragmentation spectra were identified using databases compiled of six reading frame translations of the complete genome^2–5, gene predictions⁶, expressed sequence tags (ESTs)^{5, 7–9}, homologous sequences^10–12, or exon graph models^{13, 14}, instead of reference protein databases as UniProtKB¹⁵, RefSeq¹⁶ or Ensembl¹⁷. Currently, NGS techniques are becoming more widespread and inexpensive and are regularly carried out in parallel with matching MS-based proteomics experiments. Adding extra information gathered from genomics, transcriptomics or translatomics data, respectively based on genome sequencing, RNAseq and ribosome profiling (RIBOseq), results in a more comprehensive search space for MS/MS identification. Different types of information can be obtained from these different NGS-based technologies to aid and refine the MS-based peptide and protein identification process: transcript abundance, translation efficiency, translation initiation site location, somatic versus germ line mutations, splice variation and delineation of novel coding regions ^18–21.

To help unravel the proteome complexity, search engines scan these custom protein databases, trying to identify alternative proteoforms²². Also, these custom searches can result in the identification of tumor-specific peptides^{23, 24} in onco-proteogenomics studies. The field of proteogenomics is rapidly expanding²⁵, mainly because of the advent of new sequencing techniques and the increased sensitivity and throughput of recent MS-based proteomics. An excellent review on proteogenomics concepts and applications thereof is available²⁶. In this review, we focused on the expanding toolset that is being made available to successfully analyze the merged NGS and MS datasets in proteogenomics experiments.

2. Proteogenomics goals

2.1. Aid Genome Annotation

Nowadays, due to the exponential growth of sequencing technologies, it is becoming straightforward to draft complete genomes of non-model organism helping to interpret MS/MS data²⁷ in a cost-effective way. Previously, sequence information (genome and cDNA) was unavailable for these non-model organisms, resulting in incomplete or missing protein sequences. Only homology-based or de novo algorithms could be employed to identify peptides from fragmentation spectra. These so-called homology-based algorithms allow for sequence-similarity searches^28–30 against homologues protein sequences. A plethora of proteogenomics studies were successful in plants^{2, 10–12, 31–33}, within blue technology (review by Hartmann et. al.³⁴), in prokaryotes^35–37, viruses³⁸ and within environmental microbiology (review by Armengaud et. al.³⁹) using one or both of the following two rationales: (1) homology searches using the evolutionary closest sequence (genome, cDNA, EST) template or (2) searches against an NGS-generated species-specific sequence templates.

2.2. Unravel Proteome Complexity (Identify Proteoforms)

Extra information gained from sequencing based technologies can be used to build custom, comprehensive protein databases for MS/MS spectra identification. Several types of peptides can thus be additionally identified. Figure 1 gives an overview of the different classes of peptides that can be identified in proteogenomics studies. The greater part of enzymatically cleaved peptides (e.g. using trypsin) will map to annotated protein-coding regions. The majority thereof will reside within one exonic region, but a minority can also overlap annotated transcript splice junction sites. In silico translated sequences of alternative splice isoforms, identified based on transcriptomics data (RNAseq), can be complemented to the protein sequence search DB, enabling the identification of novel splice proteoforms⁴⁰ and cross-junction peptides (covering an annotated exon-intron boundary) using MS-based techniques. Also, peptides can map to untranslated (5’UTR and 3’UTR) or intronic regions, or can point to out-of-frame translation products. Peptides starting in a 5’UTR region can give rise to upstream open reading frame (uORF) translation products or N-terminal extended proteoforms. Peptides in the 3’UTR could on the other hand point to read-through events. In some rare cases, reverse-strand peptides can also be identified².

Figure 1.

Classes of peptides identified in proteogenomics. A. A division op proteo-genomics peptide types can be made based on the genomics region where these map. The majority of enzymatically cleaved peptides map to coding genic locations (intragenic), whereas a small amount also maps to non-coding RNA and pseudogenes or intergenic regions. Exceptionally, peptides can point to chimeric proteins (fusion products in for example oncoproteogenomics studies) or could lead to gene fusion in the case of identification of gene-fusion peptides. Of the intragenic subclass, the majority will map to one exon and a minority can overspan exon splice sites (possibly leading to alternative splice isoform identification). Proteogenomics can lead to the identification of novel peptides located in untranslated regions (5’ and 3’UTR) or intronic regions, internal out-of-frame peptides, peptides that resided at the reverse strand or single amino acid variant (SAV) peptides (introduced through genetic variation or RNA editing). Other novel findings can point to exon-intron junction (cross-junction peptides). B. Another application of proteogenomics is the study of antibody or nanobody peptides in the highly variable regions. Here a combination is made of sequencing of B-cells and mass spectrometry of the blood anti/nano-bodies after affinity selection. C. Venomics is another research field wherein proteogenomics can be extremely useful. Here a combination of RNAseq of the venom gland (of for example cone snails, spiders, snails) and matching mass spectrometry boosts the identification rate of the impressively divers arsenal of toxin peptides that mostly carry multiple post-translational modifications.

Next to the aforementioned peptide classes that map to known protein-coding gene models, another type of intragenic peptides can be discovered that map to non-coding regions as (long intergenic) non-coding RNA (lincRNA) genes⁴¹ or pseudogenes^{6, 42, 43}. Moreover, peptides can map to intergenic chromosome positions, located between known genic regions. These peptides can point to translation of small open reading frames (sORFs, see below) or unannotated regular coding sequences. In some rare cases, identified peptides can result in gene joining, by identifying longer full-length open reading frames. These are located within one exon, but can also overspan annotated splice junction sites⁴⁴. Additional single nucleotide variation (SNV), insertion or deletion information, obtained from sequencing data or publicly available databases or RNA-editing sites^{45, 46} can also result in the identification of single amino acid variants (SAVs)⁴⁷. In onco-proteogenomics studies^{21, 23, 24}, usage of tumor-specific protein sequences including SNV data obtained from genomics and/or transcriptomics data regularly results in the identification of aberrant SAVs²¹. Peptides that point to fusion proteins (chimeric peptides)⁴⁸ are also regularly investigated in cancer proteogeomics studies.

Recently, a novel translatomics technique was devised, called ribosome profiling^49–52. By sequencing ribosome-protected mRNA fragments, this method charts a genome-wide protein synthesis profile. Furthermore, (alternative) translation initiation sites (TIS) can be accurately predicted by exploiting the abilities of antibiotics, such as harringtonine⁵³ or lactimidomycin⁵⁴ that stall ribosomes at sites of translation initiation. Incorporation of this information into the protein sequence database enables (alternative) TIS validation by means of matching MS-based proteomics studies^{19, 55–57}. Especially in combination with N-terminal COFRADIC (COmbined FRActional Diagonal Chromatography)^{58, 59}, that isolates amino terminal peptides, this proved to be very successful. RIBOseq is also very useful in pinpointing translation synthesis of small open reading frames^60–62. As already mentioned, these can be located in either intergenic or non-coding genic regions such as (li)ncRNA or in intronic and exonic regions of annotated genes (out-of-frame translation). Notable is that previous efforts^{63, 64} combining peptidomics and massive parallel RNA sequencing have also resulted in the identification of short open reading frame (sORF)-encoded polypeptides (SEPs) in human cells located both in coding genes (5’UTR, out-of-frame CDS, 3’UTR and antisense) and non-coding RNA.

2.3. Antibody Sequencing

The sequences of antibodies are produced by seemingly random recombination of variable (V), diversity (D) and joining (J) genes followed by somatic hypermutation. Therefore, space of potential amino acid sequences of antibodies that can be generated from a given genome is prohibitively large. Because no good sequence database was available, traditionally the only viable approach for sequencing antibodies with mass spectrometry was de novo sequencing which often involve manual labor, is time consuming, and require high-quality spectra with no missing information. However, targeted sequencing of the variable part of the antibody transcripts can provide a database that can be searched automatically and increases the success rate of identification. The combination of next generation sequencing of the transcripts that attempts at providing the sequences for the entire antibody repertoire of the individual, and affinity mass spectrometry of the circulating antibodies that bind a specific antigen is a powerful proteogenomic method to sequence antibodies. This method has been applied to characterizing circulating human antibodies in response to infection ^{65, 66}, and to production of nanobodies as research reagents based on llama single chain antibodies ⁶⁷.

2.4. Immunogenic Peptides

The application of proteogenomics approaches in the field of cancer immunology is becoming more widespread^{68, 69}. Several studies describe the application of sequencing (whole exome and/or RNAseq) in combination with mass spectrometry to identify mutated major-histocompatibility complex class 1 (MHC1) presented peptides^{70, 71}. These somatic mutations are contained in immunogenic peptides that can serve as T-cell epitopes if presented on the MHC1 molecules as they are recognized by the adaptive immune system. As more MS-based efforts are focusing on the identification and quantification of human leukocyte antigen (HLA) associated peptides^{72, 73}, the integration with tumor-specific sequencing information becomes achievable.

2.5. Venomics

Another goal of proteogenomics is the identification of venom peptides from glands of different species (e.g. spiders⁷⁴, cone snails^75–79, snakes^{80, 81}). These venoms are mostly studied because of their putative therapeutic application. Characteristic to these peptides is that they often carry disulfide-bridges that are highly stable and resistant to proteolytic degradation, ensuring that the peptide toxins remain active after administration. Another characteristic, next to the high level of post-translational modifications (PTMs), is that the venom peptide (e.g. conopeptides from cone snails) arsenal demonstrates an impressive diversity, due to diversification mechanisms at work in the venom gland. To uncover the mechanism(s) that generate this heterogeneous pool of conotoxins with alternative cleavage sites, heterogeneous post-translational modifications and highly variable amino- and carboxy-terminal truncations, an integrated approach of NGS transcriptomics with high sensitive MS-based proteomics is necessary⁷⁵. Formerly, de novo identification strategies (see below in the peptide identification section) were mainly applied to identify these venom peptides, but nowadays, with the advent of NGS techniques it became possible to construct NGS-derived venomics search databases enabling the usage of DB-search algorithms to identify these peptides.

3. Tools, databases and pipelines for proteogenomics studies

Figure 2 sketches a proteogenomic flowchart. Useful tools, databases and pipelines that can be applied throughout the different analysis steps are mentioned. The tools are subdivided based on the role they play within the proteogenomics process. Below we describe the available tools, databases and pipelines in more detail. Moreover, Table 1 gives an overview of the publicly available databases that can be used, and Table 2 lists all available tools and pipelines together with references to related literature and their website.

Figure 2.

Comprehensive overview of proteogenomics workflow. A typical proteogenomics strategy consists of different steps. Novel peptide identifications are mostly obtained by scanning comprehensive, custom-build protein sequence databases using database search engines. The search database creation step (1. DB Creation) is very important and can hold sequences from annotated protein repositories, translated genomic and/or RNA sequences or processed sequences from sequence read archives. Furthermore specific protein-related information can be incorporated in the search space. Routinely, the (very large) custom database is filtered to help manage its size. This filtering can be pursued based on different criteria. Fragmentation spectra can be experimentally obtained or downloaded from public repositories. The upper green box, connected to the arrow, gives an overview of the different proteogenomics tools reported to create such custom databases. (2. MS/MS data): PRIDE, PeptideAtlas, Massive, ProteomicsDB, Chorus, CPTAC all hold MS/MS data of MS-based proteomics experiments on different cell/tissue types and species. (3. Peptide identification): As mentioned, the identification is mostly obtained using database searching (multiple search algorithms are available). Other so-called hybrid tag-based or de novo methods have also been successfully applied in many proteogenomics experiments. (4. Validation & interpretation): After the identification step, validation and interpretation of the PSM/peptide/proteins remains indispensable, using appropriate statistical significance estimation (FDR/PEP calculation). Further global annotation analysis based on gene ontology or protein interaction is also routinely performed. (5. Mapping & visualization): The vast amount of multi-omics data can be further mapped and visualized in a genome-centric way, many tools are available for both the mapping and visualization step. Also, further integrative visualization of protein interaction networks based on this cross-omics data is also possible using several tools as Circos and Cytoscape. The lower green box, connected to the arrow, gives an overview of reported pipelines or solutions that combine the different necessary steps in a complete proteogenomics workflow.

Table 1.

Data Resources for Proteogenomics

Sequence resources - General
UniProtKB	www.uniprot.org	REF	Comprehensive, high-quality and freely accessible resource of protein sequence and functional information
RefSeq	www.ncbi.nlm.nih.gov/refseq		Comprehensive, integrated, non-redundant, well-annotated set of reference sequences including genomic, transcript, and protein.
Ensembl	www.ensembl.org		Genome databases for vertebrates and other eukaryotic species
Sequence resources - Specific
TISdb	tisdb.human.cornell.edu		Source of information for mRNA alternative translation
ChimerDB	ercsb.ewha.ac.kr/fusiongene		Knowledgebase for fusion sequences
dbCRID	dbcrid.biolead.org		Database of chromosomal rearrangements in human diseases
ChiTaRS	chitars.bioinfo.cnio.es		Database of the chimeric transcripts and RNA-seq data
CanProVar	bioinfo.vanderbilt.edu/canprovar	PMID: 21389108	Human Cancer Proteome Variation Database
dbSNP	www.ncbi.nlm.nih.gov/SNP		Database of single nucleotide polymorphisms
COSMIC	cancer.sanger.ac.uk		Catalogue Of Somatic Mutations In Cancer
HAltORF	www.roucoulab.com/haltorf		Resource for the investigation of alternative Open Reading Frames (ORFs) in human mRNA sequences
UTRdb	utrdb.ba.itb.cnr.it	PMID: 19880380	Curated database of 5' and 3' untranslated sequences of eukaryotic mRNAs, derived from several sources of primary data
uORFdb	cbdm.mdc-berlin.de/tools/uorfdb	PMID: 24163100	Comprehensive literature database on eukaryotic uORF biology
sORFdb	www.sorfs.org		Comprehensive repository of sORFs based on ribosome profiling
dbRES, DARNED	bioinfo.au.tsinghua.edu.cn/dbRES, darned.ucc.ie	PMID: 17088288, PMID: 23074185	Databases for RNA Editing sites
Animal toxin annotation	www.uniprot.org/program/Toxins		A systematical annotation of proteins secreted in animal venom
LNCiPedia	www.lncipedia.org	PMID: 25378313
OMIM	www.omim.org		Catalog for Human Genes and Disorders
Public repos for proteomics data
PRIDE	www.ebi.ac.uk/pride		Centralized, standards compliant, public data repository for proteomics data (member of ProteomeXChange, www.proteomeXChange.org)
PeptideAtlas	www.peptideatlas.org		Multi-organism, publicly accessible compendium of peptides identified in a large set of tandem MS proteomics experiments (ProteomeXChange)
MassIVE	massive.ucsd.edu		Community resource to promote the global, free exchange of mass spectrometry data (ProteomeXChange)
ProteomicsDB	www.proteomicsdb.org		Public in-memory database of the provisional human proteome map
Chorus	chorusproject.org		Web application for storing, sharing, visualizing, and analyzing spectrometry files
CPTAC	https://cptac-data-portal.georgetown.edu		Clinical Proteomic Tumor Analysis Consortium data portal
Public repos for sequencing data
SRA	http://www.ncbi.nlm.nih.gov/sra	PMID: 22009675	Raw sequencing and alignment information from high-throughput sequencing platforms

Table 2.

Tools and algorithms for proteogenomics

NGS Tools - Mappers
Bowtie	bowtie-bio.sourceforge.net/bowtie2	PMID: 22388286	ultrafast and memory-efficient tool for aligning sequencing reads to long reference sequences
TopHat	ccb.jhu.edu/software/tophat/index.shtml	PMID: 23618408	fast splice junction mapper for RNA-Seq reads
STAR	code.google.com/p/rna-star	PMID: 23104886	ultrafast universal RNA-seq aligner
NGS Tools - Other
GATK	https://www.broadinstitute.org/gatk	PMID: 21478889	a framework for variation discovery
samtools mpileup	samtools.sourceforge.net/mpileup.shtml		Calling SNPs/INDELs with SAMtools
Transcript assembly tools		PMID: 24185837	Assessment of transcript reconstruction methods for RNA-seq.
Proteogenomics Pipelines and Tools
Custom database creation
PROTEOFORMER	www.biobix.be/proteoformer	PMID: 25510491	Stand-alone or Galaxy version to convert RIBOseq data to an proteinn product database for MS-based proteomics identification.
SpliceDB	proteomics.ucsd.edu/software-tools/splicedb-splice- graph-proteomics-tools/	PMID: 25263569, PMID: 23802565	construction of a compact database that contains all useful information expressed in RNA-seq reads
CustomProDB	www.bioconductor.org/packages/release/bioc/html/customProDB.html	PMID: 24058055	Generate customized protein database from NGS data, with a focus on RNA-Seq data, for proteomics search.
Quilts	openslice.fenyolab.org/cgi-bin/quilts_cgi_v2.0.pl		Creates sample specific protein sequence databased using genomic and transcriptomic information.
Sap-DB/Splice-DB/Reduce-DB	toolshed.g2.bx.psu.edu (“Proteomics” link)	PMID: 25149441	customized proteomic databases suitable for MS searching (SAP, Splice and Reduced)
MSMSpdbb	org.uib.no/prokaryotedb	PMID: 20080508	Multi-Strain Mass Spectrometry Prokaryotic DataBase Builder
Proteomics Informed by Transcriptomics	bessantlab.org/expertise/proteomics-informed-by-transcriptomics	PMID: 23142869	Use of de novo transcript assemblies from RNA-seq data for the routine creation of sample-specific proteomes (host-pathogen or metaproteomics studies)
Complete proteogenomics pipelines
Galaxy implementations	http://toolshed.g2.bx.psu.edu, http://galaxy.nbic.nl, https://usegalaxyp.org, http://galaxy.wur.nl	PMID: 25658277	Large toolbar for MS-based proteomics and NGS integration
GenoSuite	https://sourceforge.net/projects/proteogenomic	PMID: 23882027	Framework developed for proteogenomic analysis (searching, scoring, visualization)
Peppy	geneffects.com/peppy	PMID: 23614390	Software tool designed to perform every necessary task of proteogenomic searches (create genomic DB, track loci, match MS/ MS and assign confidence)
Enosi	proteomics.ucsd.edu/software-tools/enosi	PMID: 24142994	Enosi pipeline analyzes identified peptides. Enosi is developed to automize the spectrum-peptide match to be recognized more intuitive sense.
bacterial-proteogenomic-pipeline	https://github.com/mpc-bioinformatics/bacterial-proteogenomic-pipeline	25521444	The Bacterial Proteogenomic Pipeline consists of several modules, which assist in a proteogenomics analysis.
Others
sapFinder	bioconductor.org/packages/devel/bioc/html/sapFinder.html	PMID: 25053745	detection of the variant peptides based on tandem mass spectrometry (MS/MS)-based proteomics data
PGP	https://bitbucket.org/andreyto/proteogenomics	PMID: 24470574	parallel prokaryotic proteogenomics annotation pipeline for MPI clusters based on MS proteomics data
GFS		PMID: 18428684, PMID: 12518051	peptide-mass fingerprint against the theoretical translation and proteolytic digest of an entire genome (genome fingerprint scanning).
Llama Magic	gpmdb.rockefeller.edu/llama-magic-cgi/llama_magic.pl	PMID: 25362362	Tool for identification of single chain Llama antibodies by combining next generation sequencing data with mass spectrometry based affinity proteomics data
Proteogenomics Mapping and Visualization
Mapping
PMT	www.agbase.msstate.edu/tools/pgm	PMID: 21513508	Proteogenomic mapping tool: maps peptides back to their source genome
PepLine	www.grenoble.prabi.fr/protehome/software/pepline	PMID: 18348511	maps MS/MS fragmentation spectra of trypsic peptides to genomic DNA sequences (Plants)
IggyPep	www.iggypep.org	PMID: 20000637	Maps PST to the genome
PGx	pgx.fenyolab.org		Takes peptide identities and quantities as input and and maps them onto the human genome
iPiG	https://sourceforge.net/projects/ipig	PMID: 23226516	Tool for the integration of peptide identifications from mass spectrometry experiments into existing genome browser visualizations
Visualization
peptide_to_gff	toolshed.g2.bx.psu.edu/view/galaxyp/peptide_to_gff		Tool to convert peptide to gff
PG Nexus using IGV	https://github.com/IntersectAustralia/ap11_Samifier	PMID: 24152167	covisualize peptides in the context of genomes or genomic contigs, along with RNA-seq reads (Prokaryotes - Eukaryotes)
VESPA	cbb.pnnl.gov/portal/software/vespa.html	PMID: 22480257	integrates high-throughput proteomics data (peptide-centric) and transcriptomics (probe or RNA-Seq) data into a genomic context (Bacteria)
MIMOMICs	bamics2.cmbi.ru.nl/websoftware/minomics/minomics_intro.php	PMID: 19008250	allows facile and in-depth visualization of prokaryotic transcriptomic and proteomic data in conjunction with genomics data (Prokaryotes)
Protter	wlab.ethz.ch/protter	PMID: 24162465	interactive protein feature visualization and integration with experimental proteomic data
Tools for peptide identification from MS/MS
Database searching
SearchGUI	https://code.google.com/p/searchgui	PMID: 21337703	SearchGUI is a user-friendly open-source graphical user interface for configuring and running proteomics identification search engines, currently supporting X!Tandem, MS-GF+, MS Amanda, MyriMatch, Comet, Tide and OMSSA.
PeptideShaker	https://code.google.com/p/peptide-shaker	PMID: 25574629	PeptideShaker is a search engine independent platform for interpretation of proteomics identification results from multiple search engines, currently supporting X!Tandem, MS-GF+, MS Amanda, OMSSA, MyriMatch, Comet, Tide, Mascot and mzIdentML.
Tag-based, hybrid filtering/searching
GenoMS	proteomics.ucsd.edu/software-tools/genoms	PMID: 20164058	Template Proteogenomics: Sequencing Whole Proteins Using an Imperfect Database (uses Inspect (PMID: 16013882)) and PepNovo).
Spider,TagRecon	www.bioinfor.com/peaks/features/spider.html, enchurch.mc.vanderbilt.edu/bumbershoot/tagrecon	PMID: 16108090, PMID: 20131910	Mutation-tolerant search engine.
De novo sequencing, homology searching
DenovoGUI, PepNovo+, DirecTAG	denovogui.googlecode.com	PMID: 24295440, PMID: 15858974, PMID: 18630943	User-friendly and lightweight graphical user interface called DeNovoGUI for running parallelized versions of the freely available de novo sequencing software PepNovo+ and DirecTag.
UniNovo	proteomics.ucsd.edu/Software/UniNovo.html	PMID: 23766417	Universal tool for de novo peptide sequencing
MS-blast/MS-Homology	genetics.bwh.harvard.edu/msblast, prospector.ucsf.edu/prospector/html/instruct/homologyman.htm	REFS	Tools for homology searching

3.1. Custom Database Creation

Matching fragmentation spectra to peptides contained in custom databases enabling extra identifications is an important aspect of proteogenomics. As shown in Figure 2, different types of information can be used to build an increasingly comprehensive protein sequence search database. A standard database search is performed against well-annotated resources of protein sequences, mostly obtained from UniProtKB¹⁵, RefSeq¹⁶ or Ensembl¹⁷. Within traditional proteogenomics studies, databases containing the six reading frame translation of the complete genome^2–5 or translation products from EST libraries^{5, 7–9, 82} were successfully applied. Later, with the advent of next-generation sequencing technologies, this sequencing information was incorporated in the custom protein sequence database. Next to performing matching sequencing experiments (DNAseq, exomeSeq, RNAseq or RIBOseq), many publicly available datasets are also available from sequence read archives available at the National Center for Biotechnology Information (NCBI-SRA)⁸³ and the European Bioinformatics Institute (EBI-ENA)⁸⁴. Before this sequencing information can be used, the reads that result from such an NGS experiment first need to be mapped to a reference genome. Dependent on the type of sequencing a (non) splice-aware aligner is used. Bowtie⁸⁵ or BWA⁸⁶ is the most commonly applied mapper for aligning short reads to a long reference genome or transcript sequence base. For RNAseq data though splice aware mappers (TopHat⁸⁷, STAR⁸⁸, BWA) can be used to map the reads directly to a reference genome, reporting the splice junctions with above-threshold read coverage. This threshold can be set more stringent for discovery of novel splice junctions or more tolerant for annotated splice sites. Custom databases compiled of (1) peptides that overlap (novel) exon-exon boundaries or (2) (novel) complete translation sequences using the transcript construction (e.g. based on tools as Cufflinks⁸⁹ or others described in ref. ⁹⁰) have already proven to be successful in different (onco-)proteogenomics studies^{23, 24}. Sequence variation or indels, extractable information from NGS datasets, can give rise to identification of SAVs. Most commonly, tools as GATK⁹¹ and the combination of samtools⁹² and mpileup are applied to extract this extra layer of information.

The knowledge on specific sequence features is growing and is commonly deposited in specialized databases and can thus be included in the assembly of proteins. Information on (alternative) translation initiation sites (TISdb⁹³, based on ribosome profiling) or (alternative) ORFs (HaltORF⁹⁴), untranslated regions (UTRdb⁹⁵) or evidence of translated upstream ORFs (uORFdb⁹⁶) or small ORFs (www.sORFs.org) can help to include extra layers of complexity of the proteome into the protein search space and thus help the identification thereof by means of MS-based proteomics or peptidomics experiments. Cancer fusion products (available from ChimerDB⁹⁷, dbCRID⁹⁸, ChiTaRS⁹⁹) can also be incorporated into the protein sequence database in order to detect chimeric peptides in onco-proteogenomic studies^{23, 48}. Germline or somatic nucleotide variants, deposited in the NCBI dbSNP database¹⁰⁰, supplemented with cancer mutations from CanProVar¹⁰¹ or COSMIC¹⁰² can give rise to identification of SAVs or peptides with amino acid deletion or insertions. RNA-editing events (derived from databases as dbRES⁴⁶ or DARNED⁴⁵) or calculated based on DNAseq and/or RNAseq using recently developed tools such as REDItools¹⁰³ and GIREMI¹⁰⁴ can introduce another type of genomic variation into the protein sequence database.

Information on particular classes of sequences can also be obtained for focused searches: (long intergenic) non-coding RNA can be extracted from LNCiPedia¹⁰⁵ or NONCODE¹⁰⁶ to assess whether coding sORFs can be identified, catalogs of genes related to genetic disorders can be distilled from OMIM¹⁰⁷, and protein sequences secreted in animal venoms can be obtained from the Uniprot systematical toxin annotation program (www.uniprot.org/program/Toxins).

As we pointed out, multiple layers of information can be incorporated into the protein database. This sometimes results in an explosion of the database size introducing difficulties during the peptide-to-spectrum matching (PSM). With an increasing database size, the chance that the best scoring match is incorrect also increases. In other words, distinguish the good from the bad positive matches becomes more difficult. Tools as Percolator^{108, 109} or MS2PIP¹¹⁰ are available to help boost the rate of confident peptide identifications from a collection of tandem mass spectra, using semi-supervised machine learning. Another strategy to reduce this phenomenon of increased false positive identifications is keeping the database size within limits. Several ways of filtering or redundancy removal, based on different metrics have been described. Whereas traditional proteogenomics used overlap with ESTs or gene expression as filters, RNAseq based proteogenomics studies use read coverage as a metric for expression as a cutoff (RIBOseq^{19, 55–57} or RNAseq^{18, 21, 111, 112}). To manage the size of the search database, Woo et al. ²¹ proposed a splice graph approach to compile size-limited multiple RNAseq alignment databases. Other studies use annotation information from large-scale studies to filter (e.g. based on GENCODE¹¹³) or use a chromosome-centric approach (C-HPP: Chromosome-Centric Human Proteome project ^114–116). Sequence homology¹¹ or sequence tag overlap^{2, 14, 30, 117} are also often used to help reduce the database size. A recent method (HiRIEF LC-MS) describes the use of the high-resolution isoelectric focusing⁴³ to help reduce the protein sequence DB, enabling an unbiased proteogenomics search.

Several automated tools have been proposed in the last decade to compile custom protein sequence databases. Previously, construction of a custom translated genome sequence database, using the MSMSpdbb tool¹¹⁸, was reported to study prokaryote species. The last few years, automated pipelines to construct custom protein sequence databases based on NGS data are becoming available (See Figure 2, upper green box). The majority uses RNAseq alignments to construct translated amino acid sequences, taking advantage of the multiple types of information that is available from this source of NGS data: mRNA abundance, SNVs, alternative splicing, and novel coding regions. CustomProDB¹¹² and the implementations of Sheynkeman et al. (SAP-db, SPLICE-db, REDUCED-db¹¹¹) build separated databases based on respectively variant, (alternative) splicing and mRNA expression information. Other tools try to combine all gathered knowledge within one search database using one (Quilts) or multiple (SpliceDB²¹) RNAseq alignments. PROTEOFORMER¹⁹, on the other hand uses RIBOseq data as input. Next to the aforementioned info (abundance, SNVs, splicing, novel transcripts) derived from RNAseq based experiments; RIBOseq data also enables the identification of translation initiation sites. Needless to say that this is extremely useful in delineating the true coding sequence and managing the search database size within proteogenomics studies, since only a one-frame translation is necessary from the RIBOseq-covered transcript.

3.2. Peptide Identification

Routinely, a database search engine is used to match fragmentation spectra to peptides in the database resulting in peptide identifications. Similar to regular proteomics studies, this identification strategy is also mostly applied in the proteogenomics research field. Multiple database search tools are available and nowadays results from multiple DB searches are combined to boost the identification rate. SearchGUI¹¹⁹ is a user-friendly open-source graphical user interface for configuring and running proteomics identification search engines, currently supporting X!Tandem¹²⁰, MS-GF+¹²¹, MS Amanda¹²², MyriMatch¹²³, Comet¹²⁴, Tide¹²⁵ and OMSSA¹²⁶. Next to DB-searching using different search algorithms, other hybrid approaches (e.g. InsPecT¹¹⁷, GenoMS³⁰) have been devised, filtering the search database based on peptide sequence tags (PST) derived from the actual fragmentation spectrum. This tag-based filtering allows reducing the search space, which can be very efficient and useful in proteogenomics studies. Next to this PST-based filtering, other tools (e.g. Spider²⁸ and TagRecon²⁹) allow to reconcile these tags against the protein database and moreover enable mutation-tolerant searching to identify unanticipated variations or posttranslational modifications. Finally, the true de novo sequencing algorithms are methods that do not make advantage of a reference protein database and only use spectral information for the peptide identification. These have been successfully applied in many proteogenomics studies, especially for non-model species where a comprehensive genome annotation is still unavailable or incomplete. Multiple tools to derive these partial sequence stretches or complete de novo sequences are available: e.g. PepNovo+¹²⁷, DirecTag¹²⁸ (both bundled in a graphical user interface called DenovoGui¹²⁹) and UniNovo¹³⁰. Special about the latter is that it can cope with all types of fragmentation (CID, ETD and/or HCD). Subsequently, this peptide de novo sequencing information can be used to perform database searches, looking for proteins containing peptides identical or homologues to the sequence (e.g. MS-BLAST¹³¹, MS-Homology¹³²).

3.3. Peptide validation and Interpretation

The increased size of a proteogenomics sequence database in comparison to a regular protein database can make the peptide validation very challenging. Either one assures that the protein database is similar-sized as routinely used reference protein database such as SwissProt, or a multistage search and validation strategy has to be worked out^{21, 26}. Building custom databases from RIBOseq data (where translation is only necessary from one reading frame) proved to be favorable as can be seen from the presented Posterior Error Probability (PEP) distribution plots of the identified tryptic peptides. The PEP distribution plot (presenting the cumulative valid peptide identifications for increasing PEP values) of the RIBOseq derived database behaves similar as the SwissProt one¹⁹ or a search database combining RIBOseq derived and annotated SwissProt protein sequences (after redundancy removal at the peptide level). Several other ways to manage the size of a proteogenomics database have been reviewed in the custom database creation section and are presented in Figure 2 – Customized DB.

Often, stringent validation cut-offs are used to cope with this increased DB size. The flip side of the coin is that the false negative rate will increase by minimizing the false positive rate in this way. Performing a multi-stage strategy^{21, 26, 133} has been proposed to tackle this issue. Jagtap et al.¹³³ proposed a two-step method wherein matches derived from a primary search against the large database are used to create a smaller subset database. The second search was performed against a target-decoy version of this subset database merged with a host database. A two-stage false discovery rate (FDR) strategy^{21, 26}, as opposed to a combined FDR calculation, was benchmarked by Woo et al.²¹ and proved to be advantageous using an RNA-seq derived search space built using the graph-based approach. The difference in database size (between known and NGS-derived sequences) makes that FDR cut-offs for the different PSM groups (known versus novel) need to be analyzed separately. If not, the FDR threshold for the known peptides will increase, resulting in a reduced number of known identifications, whereas the FDR cut-off for novel peptides will decrease, resulting in a higher number of false positive novel peptides.

PeptideShaker¹³⁴ is a relatively new tool to interpret MS-based proteomics identification results from multiple search engines. The PeptideShaker platform allows the combined FDR/PEP estimation approach and also enables the multi-pass strategy. The tool has a built-in function enabling the export of non-validated PSMs from a first search against for example a reference protein sequence database (SwissProt). In a follow-up analysis, the fragmentation spectra can be matched against a database of novel NGS-derived sequences. In fact, a target-decoy strategy for FDR calculation could be performed for each class of novel peptides as proposed by Nesvizhskii²⁶. Moreover, PeptideShaker allows for further separation of the PSMs based on charge or modification status if statistical significance is ensured, thus allowing a more sophisticated FDR calculation and increasing the sensitivity of the process. Further top-down validation by means of protein interaction networks (STRING¹³⁵) or gene ontology (DAVID¹³⁶, QuickGO¹³⁷) is also possible within the tool. The described multi-pass PSM confidence calculation for the different classes of novel peptides can also be incorporated in other tools such as PeptideProphet¹³⁸ within the Trans-Proteomics Pipeline¹³⁹. Next to the commonly used target-decoy strategy described above, novel ways to validate peptides have been described. The NOKOI strategy¹⁴⁰ no longer requires a decoy database and claims to be very suitable for proteogenomics approaches where very large sequence databases need to be scanned.

We already mentioned that expression information (derived from RNAseq data) could be applied to set a lower detection limit for matching follow-up MS-based proteomics measurements¹¹¹, thus incorporating transcript abundance measures within the MS database search process. These transcript abundances values have furthermore been used to also improve the sensitivity of protein identification. A study by Shanmugam et al. presented a method that uses the Global Proteome Machine Database (GPMD¹⁴¹) identification frequencies in combination with RNAseq transcript abundances to adjust the confidence score of protein identifications²⁰.

3.4. Peptide mapping and visualization

After the identification and validation of the peptides, the peptides need to be mapped to their genomic location allowing detailed inspection aiding gene or genome (re)annotation and/or covisualization with aligned RNAseq or RIBOseq data. A handful of tools are available that allow this peptide mapping in a genome-centric fashion. Some of these (e.g. Proteogenomic Mapping Tool¹⁴², PepLine¹⁴³ or IggyPep¹⁴⁴) use string-matching algorithms to map the peptide sequence to the 6 reading frame translation of the complete genome, while others (e.g. PGx or the iPiG mapper¹⁴⁵) use extra gene annotation information to speed up this mapping process. Other bioinformatics tools enable the visualization of the peptide information in the context of genomes. The Visual Exploration and Statistics to Promote Annotation (VESPA¹⁴⁶) and MINOMICS¹⁴⁷ tools are stand-alone visualization solutions that focus on the annotation of prokaryotic genomes through the integration of MS-based proteomics and NGS data in a genome-centric way. Other tools, available from the Galaxy web interface¹⁴⁸, allow the visualization of both prokaryotic and eukaryotic peptide data and facilitate the integration with other proteogenomics applications: PG Nexus using the Integrative Genome Viewer (IGV¹⁴⁹) or the peptide_to_gff tool that converts peptide data into generic feature format that can be uploaded and visualized in a genome browser environment. Yet another tool, called Protter¹⁵⁰, allows the visualization of proteoforms and interactive integration of annotated and predicted sequence features together with experimental proteomic evidence.

A standard file format holding base level PSM information and genome mapping coordinates could be very useful for future proteogenomics studies. Previously, several initiatives based on the UCSC BED format were taken to map specific proteogenomics results (e.g. from ENCODE or CPTAC project) or PeptideAtlas peptide information. Recently, the proBAM format was introduced and described, holding this base-level PSM info. A bioconductor R-package (proBAMr) is also made available converting MS-based proteomics results (in pepXML format) to this proBAM format, enabling genome-centric visualization. Higher level (e.g. peptide or ‘novel’ event) information can be extracted from this proBAM file and converted to a file format of choice (e.g. BED, BedGraph, WIG).

3.5. Integrative proteogenomics pipelines

A handful of integrative proteogenomics pipelines are available that can execute complete proteogenomics analyses. PEPPY¹⁵¹ for example starts off by generating a peptide database using the 6 reading frame translation of a known (eukaryotic) genome. Next, the peptide locations are tracked and visualized and PSM are obtained using a fixed confidence threshold. All steps (decoy DB generation, DB search and PSM validation) are automatically executed ensuring a list of identifications at the desired FDR cut-off. Other complete pipelines, applying a similar methodology, but focusing on the analysis of prokaryotic genomes, are available (GenoSuite¹⁵² and bacterial-proteogenomic-pipeline¹⁵³). ENOSI, another complete proteogenomics package initially used the same strategy in creating a genome-derived search database to help genome (re)annotation. Later, a second custom splice-graph database, compactly representing many putative splice junctions, was added to allow the identification of splice peptides. This strategy was successfully applied in a Zea mays proteogenomics study². Now, ENOSI also enables the construction of a protein database from mapped RNAseq reads, again helping to identify splice peptides. Furthermore, the pipeline still allows for the 6RF genomic translation. Peptides are identified using MS-GF+¹²¹, matching fragmentation spectra to these custom databases. Clusters of co-located peptides (i.e. events) can give rise to certain discoveries including 'novel genes, 'novel exons', novel splice junctions', or 'gene extension'. The newest kid on the block is called PGTools¹⁵⁴: a software suite for (onco)proteogenomics data analysis and visualization. This open-source software suite also completes a full proteogenomics circle and contains the several different steps: generation of RNAseq derived search databases (next to other available databases derived from Ensembl), DB searching (currently MS-GF+, X! Tandem and Comet are supported), FDR calculation (based on Percolator), grouping and annotation of proteins, identification of cancer associated proteins and visualization based on Circos¹⁵⁵, the UCSC genome browser and/or the Integrative Genomics Viewer (IGV¹⁵⁶).

ProteoAnnotator¹⁵⁷ is another open source proteogenomics annotation software designed to aid genome (re-)annotation, enabling the integration of MS-based proteomic evidence into genome databases. The goal is to use MS data for confirming that predicted gene models are translated into protein, and to verify the correctness of existing gene models. It is also usable to examine whether supporting evidence for alternative gene prediction at particular loci is present. The pipeline implements the Proteomics Standards Initiative mzIdentML standard for each analysis stage and includes steps to combine multiple search engine databases, to perform peptide-level statistics and to score grouped protein identifications matched to a given genomic locus to validate that updates to the official gene models are statistically sound.

Many tools in the genomics, transcriptomics or proteomics research field are script-based or have a command line interface. This enables researchers to create custom pipelines and workflows to perform integrative analysis of raw multi-omics data. Several frameworks (also called workflow management systems) are suitable to perform these tasks and moreover meet several necessary requirements as the need to be flexible, scalable, shareable, sustainable and complete¹⁵⁸: Taverna¹⁵⁹, Galaxy¹⁴⁸, KNIME¹⁶⁰ or Yabi¹⁶¹.

The Galaxy web platform shows to be the most practical choice for building proteogenomics workflows. Galaxy has been used within the NGS research community for almost a decade and has become the most established workflow solution. A plethora of tools for genomics and transcriptomics data manipulation and analysis has been set available over the last 10 years: e.g. for the processing and alignment of high-throughput sequencing data. Moreover, many proteomics tools have been released in the Galaxy workflow environment (a listing is provided at http://toolshed.g2.bx.psu.edu/ within the 'Proteomics' category). Complete proteogenomics pipelines can thus be built: (i) creation of custom protein databases based on RNAseq data (using the SAP-, SPLICE-, REDUCED-db tools¹¹¹ or sapFinder¹⁶²) or RIBOseq data (based on the PROTEOFORMER tool¹⁹), followed by (ii) peptide identification and validation using the SearchGUI/PeptideShaker command line interface (CLI) implementation or using the separate search algorithm implementations available. The ability to construct multi-omics workflows in the Galaxy environment is unlimited and has already been successfully applied and reported on both for proteogenomics^{43, 163} and metaproteogenomic^{133, 164} studies.

3.6. Antibody sequencing tools

Llama Magic⁶⁷ was developed as tool for identification of single chain Llama antibodies by combining next generation sequencing data with mass spectrometry-based affinity proteomics data, but it can also be used for antibody sequencing in general. It is based on these steps: (i) creation of a database of unique tryptic peptide sequences from the next generation sequencing data; (ii) searching the database using X! Tandem; and (iii) mapping the identified peptides back to the antibody sequences and scoring the sequences. The reason for creating a unique peptide sequence database instead of just translating antibody sequences and searching those is that some parts of these antibodies are identical and therefore after digestion would be highly redundant (with some peptides represented >10⁵ times). By digesting the sequences and only searching the unique peptides, the search is much faster and most search engines could not handle such a peptide redundancy. The scoring of the sequences is based on peptide coverage of the Complementarity Determining Regions (CDR’s), the overall coverage and sequence counts.

4. Conclusions and future perspectives

An elaborate list of tools is available that is applicable in proteogenomics studies, as can be seen from this review. In the current data-driven era, bioinformatics solutions for multi-omics studies are becoming more mainstream and important. The danger lies in that current worked-out integrative solutions will no longer be available in the future or that novel algorithms and improvements in both the MS-based proteomics and NGS-based sequencing research field will not find their way into the existing pipelines, if they cease to be further maintained. Although elegant integrative solutions are available (PEPPY, ENOSI, PGTools), we think that existing workflow solutions and especially Galaxy will play an important role. This very well established workflow management solution within the bioinformatics community is ideal to serve as the (default) repository for multi-omics tools and pipelines. As already mentioned, Galaxy is supported for almost a decade by a large community of researchers in the NGS domain. Over the last few years a growing interest also goes to implementing MS-based proteomics tools into Galaxy, facilitating the many proteogenomics applications listed in this review. As a large community drives this effort, the accessibility, maintenance and future update of the implemented tools in a version-based manner can also be better guaranteed. Notable is that a lot of integration is still deemed necessary to address multi-omics studies combining proteomics, transcriptomics and possibly also metabolomics and cheminformatics information.

Visualization of cross-omics data remains very important and challenging. Bundling proteomics and genomics data in a genome-centric way using different flavors of genome browsers (UCSC, Ensembl, IGV) is possible, allowing gene-level inspection of the different layers of information. This (multiple) gene-level inspection can also be done using Circos¹⁵⁵, another elegant visualization tool that allows researchers to present their multilayer data in a circular fashion. However, the main advantage of looking into multi-omics data is that it allows the inspection of interactions within complex networks over multiple layers of information. Combinations of perturbation or regulation of different nodes in an interaction network affecting the same pathway over the several layers of integrated data (e.g. transcriptomics and proteomics) need to be visualized. Cytoscape¹⁶⁵ could help in this effort and basic integration with e.g. Galaxy is available through the GenomeSpace (www.genomespace.org) implementation (a tool that brings together diverse computational tools in one place).

A point of much debate in proteogenomics studies is the FDR estimation and the confidence of the (novel) peptide identifications. In an ideal situation the scoring method can always discriminate the good from the bad positive matches even if the search space reaches the theoretical upper level of all peptide sequences possible. In a realistic situation however, this becomes increasingly difficult if more and more sequence material (based on e.g. variation) is included in the search space. A general strategy for FDR analysis has been outlined in by Nesvizhskii ²⁶ that proposes controlling for the FDR separately for peptides matching the reference database and novel proteogenomics findings. Several methods based on machine learning techniques (e.g. Percolator, MS2PIP) have been devised to perform a rescoring to increase the number of confident identifications. These methods can already aid peptide identification in proteogenomics but all the same, the ideal situation is far from reached. Until then, a multistage search strategy on successively the reference protein database and a custom genomics derived protein database can be practiced to improve the sensitivity of the peptide identification process in a proteogenomics experiment. Furthermore, this strategy imposes a separate FDR estimation for the group of known and novel peptide identifications. As mentioned before, the PeptideShaker platform can elegantly deal with this multistage strategy, even allowing further separation of the PSMs based on charge or modification status if statistical significance is ensured.

Novel proteogenomics findings, so-called ‘events’ (e.g. alternative TISs, novel proteoforms, insertions/deletions/mutations, translated UTRs, frame-shift or reverse strand translations within a known protein coding region or actual novel protein-coding loci) need to be reported and deposited in existing public protein repositories such as UniProtKB or RefSeq and genome repositories such as Ensembl or UCSC. Off course, this can only be the case for ‘events’ that have been validated using robust statistical significance estimation. Integration and visualization of novel ‘events’ from proteogenomics in genome-centric interfaces can be pursued using the track hubs system, which has also been implemented to integrate the enormous amount of information from the ENCODE project¹⁶⁶. Surely, it would be very useful and even necessary to draft community standards for this NGS-MS integration, describing what the minimum amount of information needs to be to comprehensively annotate this specific proteomics information flow. Integration with existing protein and genome resources need to be further established, on the other hand, specific, C-HPP project based dedicated web browsers are already made available to view and analyze human proteogenomics data (The Proteome Browser¹⁶⁷,CAPER¹⁶⁸, GenomewidePDB¹⁶⁹), allowing the detection of novel peptides or identification of SAVs and exon-skipping events amongst others.

In summary, proteogenomics is still a relatively new and rapidly evolving field with diverse new methods and tools being proposed, and there is also a lively discussion in the field about best practices. In several research areas a proteogenomics approach is clearly useful, for example in its utility for antibody sequencing and genome (re-)annotation. In cancer immunotherapy (i.e. immunopeptide identification) the use of proteogenomics is becoming more critical, but its usefulness for tumor analysis is still an open question. In the next few years we should see proteogenomics maturing and its application and data analysis will become more standardized.

Abbreviations

CDR

Complementarity Determining Region

CDS

Coding Sequence

CID

Collision Induced Dissociation

CLI

Command Line Interface

COFRADIC

COmbined FRActional Diagonal Chromatography

DB

Database

EBI-ENA

European Bioinformatics Institute – European Nucleotide Archive

EST

Expressed Sequence Tag

ETD

Electron Transfer Dissociation

FDR

False Discovery Rate

HARR

Harringtonine

HCD

Higher-energy Collision Dissociation

HLA

Human Leukocyte 1ntigen

lincRNA

long intergenic non-coding RNA

LTM

Lactimidomycin

MS

Mass Spectrometry

MHC1

Major-Histocompatibility Complex class 1

NCBI-SRA

National Center for Biotechnology Information – Sequence Read Archive

NGS

Next-Generation Sequencing

PEP

Posterior Error Probability

PSM

Peptide-to-Spectrum Match

PST

Peptide Sequence Tag

PTM

Post-Translational Modification

RF

Reading Frame

RIBOseq

Ribosome profiling

SAV

Single Amino acid Variant

SEP

sORF-encoded polypeptide

SNV

Single Nucleotide Variant

sORF

small Open Reading Frame

TIS

Translation Initiation Site

uORF

upstream Open Reading Frame

UTR

UnTranslated Region