What Proteomics Results Tell You, and How to Read Your Own

Oak
proteomics plasma olink somascan mass-spectrometry multi-omics
A translucent amber winged reptile-bird specimen on a steel mount, lit in a dark studio, with glowing particles suspended inside its body.

A proteomics result is a table of relative abundances: one row per protein (or per aptamer, or per antibody pair), one column per sample, and a number in each cell that is usually on a log scale and usually not in any physical unit. It tells you how much of each measured protein was circulating in that tube of plasma relative to the platform’s internal reference, not relative to a clinical normal range. That distinction drives everything downstream. A single proteomics run, on its own, is close to uninterpretable for an individual. The same run repeated four times over two years, with the same platform and matched draw conditions, is one of the densest physiological signals you can get on yourself.

What is in the file

Three delivery formats dominate, and they are not interchangeable.

Olink (proximity extension assay) ships a long-format CSV or parquet: SampleID, OlinkID, UniProt, Assay, Panel, NPX, LOD, QC_Warning. NPX is Normalized Protein eXpression, a log2 relative scale. A one-unit NPX difference is a doubling. NPX values are only comparable within an assay, across samples. Comparing NPX for IL-6 against NPX for CRP tells you nothing about their molar concentrations.

SomaScan (modified aptamers, SOMAmers) ships an .adat file: a header block with metadata, then a matrix of RFU (relative fluorescence units) with per-aptamer annotation rows. RFU is linear, not log, so you log-transform before anything parametric. Read the header fields for the normalization steps already applied (hybridization control normalization, median signal normalization, plate-scaling, and often adaptive normalization against a reference population). Applying your own median normalization on top of an already median-normalized matrix will quietly destroy real signal.

Mass spectrometry ships vendor .raw or .d, ideally converted to .mzML with msconvert --filter "peakPicking true 1-", plus a search-engine output. For data-independent acquisition, that is DIA-NN’s report.tsv (or report.parquet in recent versions) with PG.MaxLFQ intensities, or Spectronaut’s report. We prefer DIA-NN in library-free mode with --predictor --reanalyse --matrices --relaxed-prot-inf for plasma, because a predicted library plus match-between-runs recovers more precursors than a small project-specific DDA library, and the output matrices are already at protein-group level. MS gives you absolute-ish quantities (MaxLFQ intensities are proportional to abundance across samples) and, unlike affinity platforms, you can inspect the underlying peptides.

Whichever you get, the analysis-ready object is the same: a proteins × samples matrix, log2-transformed, plus a sample metadata table. Build that first, keep the raw file untouched, and version the transformation script.

Platform determines what “the protein is up” means

Affinity platforms measure binding, not protein. If you carry a missense variant in the epitope region of a target, the antibody or aptamer may bind less well and the assay reports lower abundance for a protein that is present at normal levels. These binding-site pQTLs are common enough that any per-protein result you care about should be cross-checked against your genotype for coding variants in that gene. This is one of the concrete reasons to hold sequence and protein data in the same place: the genome tells you when the protein measurement is likely an artifact 1.

Coverage differs by an order of magnitude. Olink Explore HT measures roughly 5,400 assays, SomaScan 7k/11k measures 7,000 to 11,000 aptamers, and untargeted DIA on neat plasma typically identifies a few hundred protein groups because albumin and immunoglobulins dominate a dynamic range spanning about ten orders of magnitude. Depletion columns or nanoparticle-based enrichment push MS into the low thousands at the cost of added prep variance. Targeted affinity panels win on depth for low-abundance signaling proteins. MS wins on specificity, on the ability to see proteoforms and post-translational modifications, and on being method-transparent 2.

Analyze yourself against yourself

The dominant use of proteomics in the literature is a case/control comparison across a cohort, and most tutorials teach that workflow. For a personal profile it is the wrong frame. You have n=1. What you have instead is repeated measures, and the right first analysis is a within-person longitudinal one.

Concretely:

  1. Log2-transform (skip if NPX, already log2).
  2. Drop assays where more than ~30% of your timepoints are below LOD.
  3. Center each protein on your own median across timepoints, not on a population mean. The resulting value, in NPX units, is “how far is this protein from my own baseline.”
  4. Compute a per-protein within-person standard deviation from at least three or four timepoints. That number is your noise floor for that analyte, and it varies enormously: some acute-phase proteins swing 2 NPX between draws, some structural proteins sit inside 0.1.
  5. Flag movements exceeding roughly two within-person SDs, then check whether the analyte moved together with its pathway neighbors. Correlated movement across a known module is far more credible than a single outlier assay.

Population reference ranges are still useful as a sanity check, but they are wide, and your own trajectory is more sensitive than a percentile for detecting change.

Missing values and the limit of detection

Proteomics missingness is mostly not random. A value is absent because the protein was below detection, which is informative. Mean-imputation or k-nearest-neighbor imputation will pull those values up toward the middle of the distribution and manufacture fake signal.

Our default: for affinity platforms, keep values below LOD as-is if the vendor reports an extrapolated number (Olink does), but carry the LOD flag through and never report a result driven by sub-LOD points. For MS, treat missingness as left-censored and impute from a downshifted normal (the classic Perseus approach, width 0.3, downshift 1.8 SD) only when you need a complete matrix for PCA, and re-run the key comparison on complete cases to confirm the finding survives.

Statistics that survive contact with a second dataset

With thousands of features and a handful of samples, p < 0.05 is noise. Use Benjamini-Hochberg FDR at q < 0.05 as a minimum, and expect that on a personal dataset almost nothing survives. That is the correct answer, not a failure of the analysis.

Feature selection on proteomics data is its own literature and the failure modes are well-documented: selection instability across resampling, filter methods that ignore correlation structure between proteins, and wrapper methods that overfit when features vastly outnumber samples 3. If you are going to fit a model, fit it with nested cross-validation and report the variance of selected features across folds, not just the winning set. Tools built for this, like AlphaPeptStats-adjacent ML workflows, exist partly to make the selection process inspectable rather than a single reported AUC 4.

Where protein data earns its place is as a complement to sequence and transcript data. Protein abundance is a poor proxy for mRNA abundance, and pan-cancer modeling work has shown that adding global proteomics improves prediction over transcriptomics alone 5. The general finding holds across settings: proteins carry information the other layers do not 6. Interpreting protein measurements without the genomic and transcriptional context around them leaves most of that information unused 1.

Preanalytical variables that will mislead you

  • Hemolysis. Red-cell lysis dumps hemoglobin, LDH, and peroxiredoxin-2 into plasma and shifts dozens of assays. Ask for the hemolysis index. A visibly pink sample is not usable for a longitudinal comparison.
  • Serum vs plasma vs anticoagulant. EDTA, citrate, and heparin plasma give different results. Pick one and never switch.
  • Time of day and fasting state. Cortisol-linked and metabolic proteins have real diurnal structure. Draw at the same hour.
  • Freeze-thaw cycles. Aliquot on receipt. Each thaw degrades a nonrandom subset of analytes.
  • Batch. If your timepoints run on different plates or different LC columns, batch dominates biology. Ship a bridging aliquot from a previous draw with each new batch and correct against it.

Nothing here is a diagnosis. Plasma proteomics is not a clinical test, the assays are research-use, and an unexpected movement in an inflammatory or cardiac protein is a reason to talk to a physician with the actual clinical labs in hand, not a reason to act on the NPX value.

Questions people also ask

What does proteomics data tell you? The relative abundance of circulating proteins at the moment of the draw: inflammatory signaling, tissue-leakage markers, complement and coagulation components, apolipoproteins, growth factors, and hundreds of proteins with no established interpretation. Unlike the genome, it changes week to week, which is what makes repeated measurement worth the cost.

What is proteomics in simple terms? The genome is the list of parts. The transcriptome is which parts are being ordered. The proteome is which parts are present and working right now, and it is the layer closest to physiology 7.

What are the three main types of proteomics? Expression proteomics (how much of each protein is present), structural proteomics (protein structure and complexes), and functional proteomics (interactions and activity). Consumer and research plasma profiling is almost entirely expression proteomics. Public resources like ProteomicsDB aggregate expression-level measurements across tissues and organisms and are useful for checking whether a protein you see in plasma has a plausible tissue source 8.

How much does a proteomics analysis typically cost? Affinity panels run roughly $500 to $1,500 per sample depending on panel size and volume, with per-sample cost dropping sharply at scale. Untargeted MS on neat plasma can be cheaper per sample but requires instrument access and gives shallower coverage. The real cost in a longitudinal design is the discipline of matched draws, not the assay.

Oak builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

  1. Bing Zhang, Bernhard Kuster. Proteomics Is Not an Island: Multi-omics Integration Is the Key to Understanding Biological Systems. Molecular & Cellular Proteomics, 2019. https://doi.org/10.1074/mcp.e119.001693 ↩ ↩2

  2. Andrew T. Rajczewski, Pratik D. Jagtap, Timothy J. Griffin. An overview of technologies for MS-based proteomics-centric multi-omics. Expert Review of Proteomics, 2022. https://doi.org/10.1080/14789450.2022.2070476 ↩

  3. Marta Lualdi, Mauro Fasano. Statistical analysis of proteomics data: A review on feature selection. Journal of Proteomics, 2019. https://doi.org/10.1016/j.jprot.2018.12.004 ↩

  4. Furkan M. Torun, Sebastian Virreira Winter, Sophia Doll, et al. Transparent Exploration of Machine Learning for Biomarker Discovery from Proteomics and Omics Data. Journal of Proteome Research, 2022. https://doi.org/10.1021/acs.jproteome.2c00473 ↩

  5. Mehreen Ali, Suleiman A Khan, Krister Wennerberg, et al. Global proteomics profiling improves drug sensitivity prediction: results from a multi-omics, pan-cancer modeling approach. Bioinformatics, 2017. https://doi.org/10.1093/bioinformatics/btx766 ↩

  6. Magdalena Zapalska-Sozoniuk, Lukasz Chrobak, Krzysztof Kowalczyk, et al. Is it useful to use several “omics” for obtaining valuable results?. Molecular Biology Reports, 2019. https://doi.org/10.1007/s11033-019-04793-9 ↩

  7. Richard P Horgan, Louise C Kenny. ‘Omic’ technologies: genomics, transcriptomics, proteomics and metabolomics. The Obstetrician & Gynaecologist, 2011. https://doi.org/10.1576/toag.13.3.189.27672 ↩

  8. Patroklos Samaras, Tobias Schmidt, Martin Frejno, et al. ProteomicsDB: a multi-omics and multi-organism resource for life science research. Nucleic Acids Research, 2019. https://doi.org/10.1093/nar/gkz974 ↩