Strategies combining genomics and proteomics in the identification of cancer protein biomarkers have perhaps best demonstrated the utility of proteogenomics for biomarker discovery. The Cancer Genome Atlas program (TCGA) represents a rich resource for large-scale genomic data. TCGA comprises more than 30 cancer subtypes and provides data from both cancer and control tissue (Cancer, 2006; Tomczak et al., 2015). By integrating proteomics, the Clinical Proteomic Tumor Analysis Consortium (CPTAC) has sought to expand on this dataset, performing proteomic and PTM analysis on TCGA specimens. This effort has produced robust, multidimensional proteomic datasets of cancer tissue subtypes for groups seeking to conduct integrated proteogenomic analyses (Proteomics Cancer, 2007; Ellis et al., 2013a). Several studies have successfully demonstrated the utility of these datasets in uncovering candidate biomarkers (Rodriguez et al., 2021). For example, Chiou and colleagues successfully used these data to identify S100A9 and GRN as combinatorial biomarkers for early identification of hepatocellular carcinoma (HCC) from urine (Chiou and Lee, 2016). Moreoever, Gillete and colleagues leveraged the CPTAC database to perform proteogenomic characterization of lung adenocarcinoma (LUAD) and normal, adjacent tissue (Gillette et al., 2020). This analysis utilized not only proteomic and PTM data, but also whole-exome sequencing (WES), RNA-sequencing (RNAseq), and DNA methylation analysis, to identify mRNA and peptides derived from somatic mutations as biomarker candidates of LUAD driven by ALK-fusion where fusion proteins EML4-ALK with and HMBOX1-ALK were formed at transcriptome level (Table 1).

Tumor-specific somatic mutations are ideal targets for biomarker development. For example, targeted MS-based detection of mutant KRAS_p.G12V and KRAS_p.G12D proteins has proven to be a viable biomarker strategy in colorectal and pancreatic cancers (Wang et al., 2011). In addition to oncogenic mutations, tumors have also been found to contain up to 100 “passenger” mutations, many of which are translated into potentially targetable proteins (Reddy et al., 1982; Wood Laura et al., 2007; Stratton et al., 2009; Bignell et al., 2010; Bozic et al., 2010). Although many disease-associated mutations have been identified over the years, including KRAS (Demory Beckler et al., 2013), P53 (Duffy et al., 2018), and EGFR (Awasthi et al., 2018), the vast heterogeneity of mutation sites not only poses a challenge to forming effective therapies, but also makes the possibility of creating antibodies for each mutation impractical (Leonardi et al., 2012). MS-based proteogenomics is often employed to discover mutant and novel peptides that occur downstream of tumor-specific mutations and hold promise as future biomarker candidates.

Similar to cancer, there is an increasing role for biomarkers of disease characterization and patient stratification in the field of neurodegeneration (DeKosky and Marek, 2003). Despite the fact that there has been limited success in identifying true plasma or cerebrospinal fluid (CSF) biomarkers of neurodegenerative disease thus far (Carlyle et al., 2018), there has been recent, promising progress in this field, assisted by proteogenomic strategies.

To date, the most common application of biomarkers has been in the context of disease diagnosis. Monitoring the levels of native proteins has paved the way for accurate detection of breast cancer (Gam, 2012), colon cancer (Kuppusamy et al., 2017), pancreatic cancer (Duffy et al., 2010), and neurodegenerative diseases (Heywood et al., 2015). However, there is an emerging role for the implementation of mutant protein biomarkers in disease detection. Following the established role of BRAF mutations in cutaneous melanoma, which often results in the substitution of glutamic acid for valine at position 600 (BRAF _V600E ), this genetic signature and its protein products have garnered much attention as both a diagnostic and prognostic biomarker for melanoma (Capper et al., 2011; Ghossein et al., 2013; Long et al., 2013).

Biomarker panels have demonstrated utility in detecting disease with both specificity and sensitivity. In 2017, Cohen and colleagues presented a proteogenomic screening test for the detection of pancreatic ductal adenocarcinoma using a joint panel of four conventional protein biomarkers for cancer, combined with the presence of mutant KRAS circulating tumor DNA (ctDNA) from a blood draw. With 64% specificity, 99.5% sensitivity, and a demonstrated prognostic value for overall survival, this combinatorial strategy has considerable promise for earlier detection of pancreatic cancer (Cohen et al., 2017). A year later, this strategy was expanded further by CancerSEEK, implementing a panel of ctDNA, consisting of 61 amplicons spread across 16 genes, combined with 8 protein biomarkers. CancerSEEK allows for detection of breast, colorectal, esophageal, liver, lung, ovarian, pancreatic, and stomach cancers from a single blood sample with a specificity of 99% and a sensitivity between from 69%–98%, depending on the type of cancer (Cohen et al., 2018). The efforts from Cohen et al. highlight the potential of proteogenomic panels for a variety of diseases.

In addition to diagnostic and prognostic applications, biomarkers enable patient stratification, allowing for informed and individualized treatment courses. The use of large-scale data to identify “treatable traits” in patients has been a topic of intense focus (König et al., 2017), as conventional classifications based on generalized markers have led to misclassification and ineffective treatment of clinically and pathologically heterogeneous disorders (Nevo et al., 2016). In an attempt to expand upon the five currently implemented breast-cancer subtypes derived from a set of 50 transcriptional signatures (i.e., PAM50 markers) (Parker et al., 2009), Johansson et al. utilized an integrated proteomics analysis on tumor tissue from patients representing each of the five PAM50 subtypes. (Johansson et al., 2019). In addition to identifying proteins derived from non-coding regions as candidate immunotherapeutic targets, network analyses succeeded in stratifying known patient classifications further, proposing previously unrecognized biomarkers and subclasses to guide therapeutic development.

Two studies in lung adenocarcinoma have also highlighted the potential of applying proteogenomics in patient stratification. Chen et al. revealed 5 mutational profiles previously unidentified in LUAD in an East Asian cohort (Chen et al., 2020). The group identified protein and genetic signatures in these subtypes strongly tied to age, gender, and EGFR-mutation status, contributing important considerations for the development of disease-modifying therapies. Furthermore, integrated analyses of multi-omics data from glioblastoma (GBM) samples unveiled new immune-based subtypes, expanding on previous classifications based only on transcriptomic and genomic data (Wang et al., 2017; Wang et al., 2021). Notably, the study subdivided glioblastoma into two distinct groups, allowing for future, more in-depth mechanistic studies to reveal therapeutic vulnerabilities in these newly discovered subclasses for precision medicine (Oh et al., 2020). Leveraging genomic, transcriptomic, and proteomic data together has provided rich resources for better patient stratification, as well as the identification of potential biomarker and therapeutic targets.

Here, we propose a general MS-based proteogenomic workflow for the identification of variant protein markers in human biospecimens (Figure 2). The first step in creating customized databases capable of detecting variants in MS-based approaches is to identify disease-relevant genomic variants. Informatic tools for variant calling are widely available. The most common variants are SNV variants—commonly identified through tools such as Platypus (Rimmer et al., 2014) and Samtools (Li, 2011)—and splicing variants—which can be identified using MAJIQ (Vaquero-Garcia et al., 2016) and MapSplice (Wang et al., 2010), among other tools. Novel peptide products can be predicted from RNA-sequencing results via ECgene (Lee et al., 2006), FastDB (De La Grange et al., 2005), FANTOM3 (Carninci et al., 2005), or the ASTD (Koscielny et al., 2009). Novel protein sequences generated from in silico translation of the reference genome and/or transcriptome—e.g., via tools such as AGUSTUS (Stanke et al., 2006), GENEID (Parra et al., 2000) or EuGENE (Foissac et al., 2003)—allow for customized databases with the power to identify and validate proteins and peptides translated from antisense strands, non-coding genes, intergenic regions, and untranslated regions (UTRs) (Nesvizhskii, 2014). Once the RNA sequences of interest are identified, in silico translation tools, such as Transeq (CITE), Quilts (Ruggles et al., 2016), and GalaxyP (Sheynkman et al., 2014), can be used to predict the resulting amino acid sequence and build a custom peptide database. With this customized FASTA database, it is possible to perform searches of proteomics raw files for sequences of interest using MS search engines, such as PEAKS (Tran et al., 2019), Proteome Discoverer, and MaxQuant (Cox and Mann, 2008).

Integrated proteogenomic algorithms are also available for “one-stop” analyses, starting from variant calling to MS-spectra annotation (Table 2); however, some tools are not as popular as database search engines and have not been thoroughly validated. Beyond generating patient-specific databases, common mutations from existing databases (Table 3) can be introduced to native proteome databases. For example, Catalogue of Somatic Mutations in Cancer (COSMIC), containing somatic mutations from variety of cancer types, has been widely used for generating customized reference and identifying cancer-specific mutations (Zhu et al., 2018). Qi and colleagues utilized LNCipedia to predict lncRNAs regions and discovered lncRNA-coded neoantigens in lung adenocarcinoma (Qi et al., 2021). A key consideration in developing a proteogenomic database search strategy is the determination of an appropriate false-discovery rate (FDR). By increasing the database size through the integration of native plus variants proteome, the identified variant peptides are prone to high false positive rates from multiple comparisons. Therefore, additional targeted methods are required for validation.

Similar to NGS approaches, MS-based proteomics has rapidly advanced throughout the past two decades. Performing total RNA-seq in biofluids has proven to be technically challenging (Everaert et al., 2019). Given the low quantity and quality of RNA in biofluids, most biomarker studies focus on circulating DNA and small RNAs (Buschmann et al., 2016; Vo et al., 2019). Therefore, protein biomarkers have become the most common clinical markers in body fluids. To increase proteome coverage, various approaches have been adopted. These include a) offline fractionation to reduce sample complexity; b) high-abundant protein depletion to remove housekeeping proteins in biofluids; c) enrichment for tissue-derived extracellular vehicles (EVs) (Fiandaca et al., 2015; Mustapic et al., 2017; Heath et al., 2018); d) nanoparticle-based enrichment of low-abundant proteins and co-depletion of high-abundant proteins (Kim et al., 2018; Tiambeng et al., 2020); and e) use of multiple proteases to detect peptides not typically generated by standard trypsin cleavage (Giansanti et al., 2016).

For data acquisition in discovery proteomics, data-dependent acquisition (DDA) and data-independent acquisition (DIA) are commonly used in MS. Previous studies demonstrated that DDA and DIA acquire different groups of peptides; this could extend the pool of total peptide identification and protein coverage (Reilly et al., 2021). DDA typically generates less complex, but more specific, MS2 spectra of selected peptides; however, only the most abundant peptide precursors are selected. On the other hand, DIA is a more inclusive approach to fragment all peptide precursors, including low-abundant ones. Although DDA has been more widely applied in biomarker studies, DIA has gained traction more recently for its applications in identifying low-abundant peptides (Guo et al., 2015; Latonen et al., 2018). The increased scan speed of high-resolution MS allows DIA to use narrower isolation windows and cover a broader m/z range (e.g., 400–1,000). DIA generally provides higher confident peptides due to the longer MS2 injection time, which allows for high-resolution MS2 spectra. Database search of DIA data typically requires a spectral library generated from the respective DDA MS run; notably, recent studies demonstrate the direct application of DIA data using a protein sequence library where “pseudo-spectra” and predicted retention times of each precursor ion is generated by search engines, such as DIA-Umpire (Tsou et al., 2015), Spectronaut, and DIA-NN (Demichev et al., 2020). Emerging evidence shows DIA is the next-generation data acquisition approach for label-free proteomics.

Targeted proteomic analyses are commonly employed to validate mutant peptides discovered through DIA/DDA shotgun proteomics and to generate high-throughput MS-based assays for clinical use. Targeted approaches, including multiple reaction monitoring (MRM) and parallel reaction monitoring (PRM), align select or all MS2 transitions and retention times of in vivo peptides and their “heavy isotope” synthetic counterparts that serve as internal standards. Typically, a list of m/z ratio of the precursor ions and their daughter ions is built into the MS instrumentation method to selectively monitor targets. Furthermore, DIA is a “semi-targeted” approach, as the MS2 transitions that are used for qualification can also be visualized as PRM-like spectra in Skyline (MacLean et al., 2010) and SpectroDive. Many proof-of-concept studies have utilized targeted methods to validate variant peptides, as the “gold standard,” ultra-sensitive approach. The biomarker specificity of validated peptides should also be demonstrated in large-scale cohorts containing disease and healthy control samples. If the variant peptides are validated as specific biomarkers, scalable MS-based MRM assays can be developed to rapidly detect such biomarkers in patient samples for point-of-care diagnosis and disease subtype stratification.