Genomics vs. Proteomics: What Each One Measures and How They Differ
Genomics measures a sequence that does not change. Proteomics measures a concentration that changes by the hour. That single distinction drives nearly every practical difference between the two datasets. It shapes file formats and error models, how often you should sample, and what you are allowed to conclude from a single measurement.
Concretely, whole-genome sequencing reads the roughly 3.1 billion base pairs of your DNA and reports where you differ from a reference. The answer is essentially the same whether the blood was drawn today or in twenty years. Proteomics measures the abundance of protein molecules in a specific sample, usually plasma or serum. Those abundances respond to sleep, infection, and exercise. They also shift with fasting state and time of day. Genomics gives you a fixed set of constraints. Proteomics gives you a snapshot of state within those constraints.
What the genome measurement is
A standard personal genome starts as short reads from an Illumina instrument or long reads from Oxford Nanopore or PacBio. The raw output is FASTQ: sequence plus per-base quality scores. Reads get aligned to a reference genome, usually GRCh38 or the newer T2T-CHM13 assembly, producing a BAM or CRAM file. CRAM is what you want for storage, since reference-based compression takes a 30x genome from roughly 100 GB down to 20-30 GB.
Variant calling turns alignments into a VCF file. With DeepVariant or GATK HaplotypeCaller at 30x coverage, you should expect roughly 4 to 5 million variants relative to GRCh38, the large majority of them single-nucleotide variants already catalogued in gnomAD. The interesting fraction is small. Perhaps 10,000 to 12,000 variants alter protein coding sequence, and only a few dozen to a few hundred are rare enough and predicted damaging enough to merit a second look.
The error model here is well characterized and mostly positional. Short reads do poorly in segmental duplications, in the HLA region, and in repeat expansions. They also do poorly in structural variants larger than the read length. You can audit these failures directly. Open the BAM in IGV and look at the read pileup. Check the depth and the mapping quality, then decide whether a call is real. That kind of per-observation audit is the defining convenience of genomic data.
One consequence deserves emphasis. Because the sequence is stable, one good genome is enough. Re-sequencing yourself adds almost nothing except better coverage in hard regions. Money spent on a second genome is better spent on a second time point of something dynamic.
What the proteomics measurement is
Proteomics measures protein abundance, and it does so indirectly. Two platform families dominate. Mass spectrometry digests proteins into peptides with trypsin, separates them by liquid chromatography, and identifies them by mass and fragmentation pattern. Affinity platforms such as Olink and SomaScan use antibodies or aptamers that bind specific target proteins and report a relative signal.
The distinction matters for interpretation. A mass spectrometry run is a discovery measurement: it will find whatever peptides are abundant enough to detect, and identification depends on matching spectra against a protein sequence database. Affinity panels measure a predefined list, typically 1,500 to 7,000 targets, with good sensitivity but a dependence on the binding reagent being specific for what its label claims.
Plasma proteomics is hard for one structural reason. Protein concentrations in blood span more than ten orders of magnitude, from albumin at tens of milligrams per milliliter down to cytokines in the picogram range. Albumin and immunoglobulins alone account for the bulk of plasma protein mass. Untargeted mass spectrometry therefore sees the abundant proteins well and the interesting signalling proteins poorly, unless you deplete the top species or enrich your targets first.
Output formats reflect the two-stage nature of the analysis. Raw instrument files such as Thermo .raw or the open mzML feed a search engine such as MaxQuant, FragPipe, or DIA-NN. The search engine produces a peptide and protein quantification matrix. What you receive as a user is usually that matrix: proteins as rows, samples as columns, with normalized intensities or NPX values. Missing values are common and are not random. A protein absent from one sample often means it fell below the detection limit there. The protein was not necessarily absent from the blood.
Why the two datasets answer different questions
Genomics tells you about constraints on your biology that were set at conception. If you carry a rare loss-of-function variant in a gene, that is true regardless of physiological state, and it shifts your prior on a whole class of downstream observations. Proteomics tells you which of those constraints is currently expressed in the molecules doing work. Pandey and Mann framed this as the reason proteomics remained necessary after the genome was sequenced: gene sequence does not specify protein abundance, modification state, or localization.1
The gap between the two is larger than people expect. Alternative splicing, post-translational modification, and protein turnover all break any simple mapping from gene to protein. A gene may produce several proteoforms with different activities. Phosphorylation and glycosylation change function without changing sequence. Yates and Osterman describe the two fields as advancing in parallel because each addresses a layer the other cannot reach.2
The relationship runs in both directions. Proteogenomics uses personal sequence data to build a custom protein search database, so that mass spectra can be matched against peptides your variants encode rather than against a generic reference proteome. This recovers variant peptides that standard searches miss entirely.3 In the other direction, observed peptides serve as experimental evidence for gene annotation, confirming which open reading frames are translated and correcting predicted exon boundaries.4
How we would use each in a personal dataset
We treat the genome as a one-time structural layer and the proteome as a repeated-measures layer. Sequence once at 30x or higher, keep the CRAM and the gVCF, and revisit the annotations as ClinVar and gnomAD improve. The FASTQ and CRAM are the durable assets. Any interpretation layered on top will be reanalyzed.
For proteomics, a single time point is close to uninterpretable on its own for most analytes. There is no population reference range for most of the 5,000 proteins on a large panel, and the biological variability within one person across days is substantial. What makes proteomics useful personally is a within-person baseline: three or more time points under comparable conditions, collected in the morning and fasted, with consistent tube type and processing time.
Pre-analytical handling dominates the noise budget. Platelet activation during clotting releases proteins into serum that were not free in circulating plasma, so serum and EDTA plasma give different answers for the same person. Time from draw to centrifugation, freeze-thaw cycles, and hemolysis all shift measured values. If you compare two time points processed differently, you are measuring your logistics rather than your biology.
Analysis-wise, we work from the protein matrix in R or Python with log2-transformed intensities, and we model each protein against its own history rather than against a cohort. Plot the trajectory before running any statistics. Most apparent signal in a small personal series is a handling artifact or a single outlier run, and a plot finds it faster than a test does.
Where does interpretation stop? Both layers are measurements, and neither supports a clinical conclusion on its own. A rare variant in a disease-associated gene needs a clinical geneticist to assess penetrance and family history. An unusual inflammatory or cardiac protein pattern needs a physician who can put it next to your symptoms and standard labs. The value of the data is that it gives you and your clinician something specific to reason about. Genomics and proteomics have long been discussed as complementary inputs to nutrition and physiology research for exactly this reason.5
Questions people also ask
Is proteomics just a readout of gene expression?
No. RNA sequencing measures transcript abundance, and transcript levels correlate only moderately with protein levels, typically in the range of 0.4 to 0.6 for the same tissue. Protein abundance also depends on translation rate and degradation rate, and neither is visible in RNA data.1
Can I predict my proteome from my genome?
Only in narrow cases. Some variants have large, reproducible effects on the measured level of a single protein, often because they sit in or near the gene encoding it. For most proteins the genetic contribution is a modest fraction of the observed variance, and environment and state account for the rest.
Which should I get first if I can only do one?
Get the genome first. It is a one-time measurement with a stable answer, it never needs repeating, and it makes later proteomic and transcriptomic data more interpretable by defining which protein sequences you carry.3
Why do my two proteomics reports disagree?
Different platforms measure different things and report on different scales. An Olink NPX value and a SomaScan RFU value for the same named protein are not interchangeable, and agreement between affinity platforms for a given target is often weak. Compare within a platform, and keep tube type and processing constant.
What file formats should I insist on?
For genomics: FASTQ or CRAM, plus a VCF or gVCF with the reference build stated. For proteomics: the normalized protein-level matrix as CSV or TSV, plus the per-sample QC metrics. If mass spectrometry was used, also ask for the mzML or vendor raw files so a search can be rerun against your own variant database.
Footnotes
-
Akhilesh Pandey, Matthias Mann. Proteomics to study genes and genomes. Nature, 2000. https://doi.org/10.1038/35015709 ↩ ↩2
-
John R. Yates, Andrei L. Osterman. Introduction: Advances in Genomics and Proteomics. Chemical Reviews, 2007. https://doi.org/10.1021/cr068201u ↩
-
Gloria M. Sheynkman, Michael R. Shortreed, Anthony J. Cesnik, et al. Proteogenomics: Integrating Next-Generation Sequencing and Mass Spectrometry to Characterize Human Proteomic Variation. Annual Review of Analytical Chemistry, 2016. https://doi.org/10.1146/annurev-anchem-071015-041722 ↩ ↩2
-
C. Ansong, S. O. Purvine, J. N. Adkins, et al. Proteogenomics: needs and roles to be filled by proteomics in genome annotation. Briefings in Functional Genomics and Proteomics, 2008. https://doi.org/10.1093/bfgp/eln010 ↩
-
Daniel H.*. Genomics and proteomics: importance for the future of nutrition research. British Journal of Nutrition, 2002. https://doi.org/10.1079/bjnbjn/2002554 ↩
Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.