NPX (Normalized Protein eXpression)

NPX is a relative, log2-scaled unit of protein abundance produced by Olink’s Proximity Extension Assay, where one NPX unit corresponds to a two-fold difference in signal and the zero point is arbitrary. It is not ng/mL, and NPX for IL-6 and NPX for TNF in the same sample are not comparable to each other.

How it works

Each assay uses two antibodies against the same target protein, each carrying a short oligonucleotide. When both bind the same protein molecule, the oligos come into proximity, hybridize, and get extended into a unique amplicon. That amplicon is counted either by qPCR (Target 48, Target 96) or by DNA sequencing (Explore 384 up to Explore HT, ~5,400 proteins). The sequencing readout is why people ask about “Olink sequencing”: the sequencer is a molecule counter, not a protein sequencer. Requiring two antibodies on one molecule is what keeps cross-reactivity low enough to multiplex thousands of assays in 1–3 µL of plasma.

The raw readout is Ct (qPCR) or read count (Explore). NPX is built by subtracting an extension control (within-sample technical variation), subtracting an inter-plate control (plate effect), and then subtracting a per-assay correction factor that sets the arbitrary zero. On the qPCR panels the sign is flipped so that higher NPX means more protein. The result is log2, so a difference of 1.0 NPX is a doubling and 0.3 NPX is roughly a 23% change.

Yes, this is targeted proteomics. You get exactly the proteins on the panel you bought, with no discovery of unexpected proteoforms, and in return you get sub-pg/mL sensitivity on cytokines that mass spectrometry cannot reach in unfractionated plasma ¹.

In your own data

Olink delivers a long-format file (.parquet or .csv), one row per sample per assay. The columns that matter:

• OlinkID — the stable key. Use it, not Assay. The same gene name appears on multiple panels (IL6 sits on Inflammation and Cardiometabolic) with different OlinkIDs and different NPX scales.
• UniProt — join key to everything else.
• NPX, LOD, MissingFreq, QC_Warning, Assay_Warning, PlateID, Panel.

A workflow we would use, in R with the official OlinkAnalyze package:

df <- OlinkAnalyze::read_NPX("explore_ht.parquet")
olink_qc_plot(df)        # IQR vs median NPX per sample, flags outliers
olink_dist_plot(df)      # per-sample NPX distribution
df <- dplyr::filter(df, QC_Warning == "PASS", Assay_Warning == "PASS")

Then drop assays where MissingFreq > 0.25 in your cohort. Do not drop individual below-LOD values and impute them. Olink’s own guidance is to keep the reported NPX below LOD rather than censor, because the values still carry rank information, and censoring induces a spike at the detection floor that breaks parametric tests. For a single person’s longitudinal series, an assay that sits under LOD at every timepoint is noise, and you should remove the assay entirely rather than interpret its wobble.

Two failure modes dominate. First, cross-panel or cross-version comparison: Explore 3072 NPX and Explore HT NPX for the same protein are on different scales, and Olink’s bridging requires a set of ~40 shared samples run on both, after which you apply a median-shift per OlinkID (olink_normalization_bridge()). Without bridging, your “change over 18 months” is a reagent lot change. Second, plate layout. If your baseline draw is on plate 1 and your follow-up on plate 2, the plate effect and the biology are the same variable. Randomize or, for a personal series, freeze aliquots and run all timepoints on one plate.

For interpretation, the useful signal is usually multivariate. Protein age clocks built on plasma panels separate organ-specific aging trajectories and predict mortality and disease risk beyond chronological age ². Proteomic scores also capture disease signal that a polygenic score misses, because the proteome integrates environment and current state rather than germline risk alone ³.

Limitations

NPX is relative. You cannot say “my IL-6 is 3.1 pg/mL” from NPX without a separate absolute assay or an Olink kit that ships a standard curve. Comparison is within-assay, across samples, only.

Epitope effects are real. A missense variant in the binding region can lower measured NPX with no change in protein concentration, which shows up as a strong cis-pQTL with a single-variant signature. Proteogenomic work treats this as a recognized confounder and a signal to be checked before believing a cis association ⁴. The same machinery, applied carefully, is powerful: plasma proteomics has surfaced causal candidates via Mendelian randomization ⁵ and has yielded diagnostic leads in patients left undiagnosed after genome sequencing ⁶.

Panel content constrains the question. Inflammation panels dominate the published literature because they work well, and they distinguish disease severity strata cleanly ^{7 8}, but a normal inflammation panel is not a normal proteome.

Finally, no NPX value is a diagnosis. Out-of-range results, including ones that look dramatic across a longitudinal series, need a clinician and a validated clinical assay before they mean anything about your health.

Woolf Software 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.