Proteogenomics: concepts, applications and computational strategies

Abstract

Proteogenomics is an area of research at the interface of proteomics and genomics. In this approach, customized protein sequence databases generated using genomic and transcriptomic information are used to help identify novel peptides (not present in reference protein sequence databases) from mass spectrometry-based proteomic data; in turn, the proteomic data can be used to provide protein-level evidence of gene expression and to help refine gene models. In recent years, owing to the emergence of new sequencing technologies such as RNA-seq and dramatic improvements in the depth and throughput of mass spectrometry-based proteomics, the pace of proteogenomic research has greatly accelerated. Here I review the current state of proteogenomic methods and applications, including computational strategies for building and using customized protein sequence databases. I also draw attention to the challenge of false positive identifications in proteogenomics and provide guidelines for analyzing the data and reporting the results of proteogenomic studies.

Figure 1. Peptide and protein identification in shotgun proteomics

A) Overview of shotgun proteomics. Proteins are digested into peptides, then separated using liquid chromatography coupled online to a mass spectrometer, then analyzed by the mass spectrometer which generates tandem mass (MS/MS) spectra. B) Peptides are most commonly identified using a sequence database search approach. Traditionally, experimental MS/MS spectra are matched with theoretical spectra predicted for each peptide contained in a protein sequence database. Sequence tag-assisted database searching starts with extraction of short tags followed by database searching in which the list of candidate peptides is restricted to those peptides only that contain one of the extracted sequence tags, allowing for mutations in the sequences of candidate database peptides. Peptide sequence can also be extracted directly from the spectrum using de novo sequencing (extracted sequences can then be searched in a protein sequence database to find the exact or a homologous peptide).

Figure 2. The concept of proteogenomics

In a proteogenomics approach, genomics (DNA sequencing, expressed sequence tags (ESTs) and transcriptomics (RNA-Seq, ribosome profiling) data is used to generate customized protein sequence databases to help interpret proteomics (LC-MS/MS) data. In turn, the proteomics data provides protein-level validation of the gene expression data, as well as helping to refine gene models. The enhanced gene models can help improve protein sequence databases for traditional proteomics analysis.

Figure 3. Type of peptides identified in proteogenomics

Peptides identified by searching customized protein sequence databases are mapped on the genome. Intergenic peptides map to regions located between annotated gene models, whereas intragenic peptides map to genomic regions contained within or in close proximity to an annotated gene model. Intragenic peptides can be further categorized based on the annotation of the corresponding gene model (e.g. ‘protein-coding gene’, ‘long noncoding RNA (lncRNA) gene’, and ‘pseudogene’). The majority of peptides map to a protein coding gene, and can be divided into Exon and exon-exon junction (Junction) peptides. Novel peptides include peptides mapping to untranslated regions (3′ or 5′ UTR peptides) or Intron peptides, peptides spanning the boundary between the coding sequence region and the neighboring UTR or intron region (Exon boundary), peptides spanning un-annotated (alternative) splice junctions (Alt junction), and out of frame peptides (Alt frame).

Figure 4. Statistical assessment of peptide identifications in proteogenomics

MS/MS spectra are searched against a customized protein sequence database that includes target sequences for the organism of interest, i.e. a reference protein database and predicted protein sequences (containing novel peptides). In addition, two decoy databases (e.g. reversed sequences) of the same sizes as the target reference and predicted databases are appended to the target databases. The best database peptide match for each spectrum is selected for further analysis. Peptide identifications are classified as known or novel (for a decoy peptide the class - ‘known’ or ‘novel’ – is determined based on the class of the corresponding target sequence from which the decoy was generated). When using simple database search score based filtering, the numbers of target and decoy peptide identifications passing a certain score threshold are counted and used to estimate FDR corresponding to that threshold. FDR analysis should be done separately for known and novel peptides (class-specific FDR) due to difference in the number of known and novel sequences in the searched customized sequence database, and due to lower likelihood of correctly identifying a novel peptide than known peptide. When using more advanced methods based on computing posterior peptide probabilities, both the database search scores and the peptide class (known or novel) should be taken into consideration.