Dual antibody-oligonucleotide proximity probes hybridize, extend, and amplify — turning single protein-binding events into a quantitative qPCR/NGS readout
Every Olink experiment begins with a remarkably small input: a single microliter of undiluted plasma or serum, dispensed directly into a 96- or 384-well plate without upfront depletion of abundant proteins like albumin. This is possible because Proximity Extension Assay (PEA) chemistry is inherently high-specificity and background-suppressing, so the same crude biofluid that would overwhelm a standard sandwich ELISA can be probed for thousands of analytes in parallel from one well.
Olink panels are organized in tiers that trade multiplexing breadth against per-target sensitivity and cost:
• Olink Target 96 — 92-plex thematic panels (Inflammation, Oncology, Cardiometabolic, Neurology) read out by qPCR on the Olink Signature Q100; ideal for hypothesis-driven, high-throughput clinical studies with fast (<24h) turnaround. • Olink Target 48 / Focus — compact 21–48 protein panels for rapid, cost-efficient screening of a specific pathway. • Olink Explore 384 — three or more 384-plex panels (~3,072 proteins) read out by next-generation sequencing (NGS-PEA), used heavily in large biobank-scale discovery studies. • Olink Explore HT — the current flagship platform, quantifying 5,416 proteins from a single 1 µL sample on one NGS run, combining 384-well automation with Illumina NovaSeq 6000 sequencing.
Sample requirements are deliberately minimal: EDTA plasma is preferred over serum (avoids clotting-cascade artifacts on coagulation-pathway proteins), 1–3 freeze-thaw cycles are tolerated, and hemolysis/lipemia are flagged automatically by QC warning flags embedded in every plate.
Because a single Explore HT plate simultaneously measures 5,416 analytes across up to 88 samples (with 8 wells reserved for controls), reproducibility depends entirely on automation. Olink Flex and Agility liquid handlers pipette antibody-oligo probe cocktails into each well at femtomolar-to-picomolar concentrations, incubate at 4°C overnight (proximity binding), then hand off to the extension/amplification and library-prep stages without manual intervention.
Every plate carries: • Negative controls (buffer only) — establish assay background • Inter-plate controls (IPC) — a pooled plasma sample run on every plate, used later for bridging normalization • Sample controls — replicate patient samples to estimate technical CV
This architecture is what allows Olink data to be pooled across cohorts collected years apart at different sites — a critical requirement for biobank-scale proteomics.
The specificity of PEA does not come from any single antibody — it comes from requiring two independent recognition events to co-occur on the same protein molecule at the same time. Each antibody in a matched pair is conjugated to a distinct short single-stranded DNA oligonucleotide; only when both probes are bound to neighboring epitopes on one target does their combined local concentration become high enough for the complementary oligo tails to hybridize.
Each PEA probe pair is built from two matched monoclonal or polyclonal antibodies raised against non-overlapping epitopes of the same target protein:
• ProbeA: antibody covalently linked (via a flexible linker) to a short oligonucleotide bearing a 3′ single-stranded overhang • ProbeB: the partner antibody linked to a complementary oligonucleotide bearing a 5′ overhang
When both probes bind their epitopes on the same protein molecule, the antibody-oligo conjugates are held in constrained proximity — typically within a Förster-like effective radius of roughly 15–20 nanometers. At that distance, the local effective concentration of the two oligo tails increases by several orders of magnitude relative to bulk solution, driving specific hybridization even though the oligo sequences themselves are only 15–20 bases long and would not hybridize efficiently in free solution at assay concentrations.
This dual-recognition requirement is the single largest contributor to PEA's specificity advantage: a single stray antibody binding a homologous off-target protein produces no signal, because its solo oligo tail has no partner nearby to hybridize with. Sandwich ELISAs, by contrast, can still generate false-positive signal from any single high-affinity cross-reactive antibody.
Running thousands of probe pairs in the same well raises an obvious risk: with 5,416 different ProbeA/ProbeB combinations floating in one tube, could ProbeA from analyte #1 hybridize with ProbeB from analyte #4,000 and generate spurious signal? Olink's oligo design mitigates this through:
• Sequence orthogonality — every oligo tail across the panel is designed with minimal cross-hybridization potential, verified computationally against the full panel oligo library • Distance-dependence — hybridization efficiency falls off steeply beyond the ~20 nm proximity window, so even if two unrelated probes drift near each other in solution, transient random collisions rarely persist long enough to extend • Low probe concentration — each probe is present at picomolar levels, keeping the total oligo pool dilute enough that random bimolecular hybridization is rare relative to proximity-driven events
The net effect: measured non-specific background across a 5,416-plex Explore HT panel is typically below 2% of the assay's total dynamic range, allowing genuinely low-abundance cytokines (many circulate at low picogram/mL concentrations) to be resolved above noise.
Proximity alone does not produce signal — it produces an opportunity. Once the two oligo tails hybridize, a DNA polymerase extends each strand using the other as template, filling in a short double-stranded region that becomes a unique, protein-specific amplicon. That amplicon is then PCR-amplified and read out either by real-time qPCR (lower-plex panels) or by pooling and sequencing on a next-generation sequencer (high-plex Explore panels).
The hybridized oligo pair forms a short duplex with 3′ recessed ends pointing toward each other. A DNA polymerase added to the reaction extends both strands, using the opposing oligo as template, generating a new double-stranded DNA sequence that did not exist before the two probes met. Because this new sequence is only created when a bona fide proximity event occurs, its molecular count is a direct, quantitative proxy for the number of target protein molecules bound in that well.
The resulting amplicon has three functional segments: • A protein-identifying barcode sequence unique to that probe pair • A universal PCR priming site shared across the whole panel (enabling single-primer-pair pre-amplification of all 5,416 targets simultaneously) • For NGS panels, a sample-identifying index and unique molecular identifier (UMI) added during library preparation, allowing digital deduplication of PCR amplification bias during sequencing analysis
A universal PCR pre-amplification step (roughly 17–20 cycles) boosts total amplicon abundance before splitting the pool for either qPCR or sequencing readout, without disturbing the relative ratios between different protein barcodes.
Once amplicons exist, Olink offers two alternative quantification routes depending on panel plex:
qPCR readout (Olink Signature Q100, lower-plex Target panels): • Microfluidic chip runs thousands of nanoliter-scale real-time PCR reactions in parallel • Each analyte-specific primer pair amplifies only its matching barcode • Cycle threshold (Ct) is inversely proportional to log2 protein abundance — low Ct (~14–18) indicates a highly abundant target, high Ct (~28–30) indicates near-background • Full 96-sample × 96-analyte run completes in under 3 hours
NGS readout (Explore panels, up to 5,416-plex): • All barcoded amplicons across the whole panel and all samples on a plate are pooled into one sequencing library • Sequenced on an Illumina NovaSeq 6000, typically generating tens of millions of reads per sample • Each read is mapped back to its protein barcode and sample index; UMI collapsing removes PCR duplicate reads before counting • Read counts per protein per sample are the raw quantitative unit, later converted to NPX
Both readouts span roughly five orders of magnitude of dynamic range per analyte, sufficient to capture cytokines from low pg/mL up to structural or highly abundant plasma proteins in the same run.
Raw Ct values or NGS read counts are not directly comparable across plates, runs, or panels — they are influenced by pre-amplification efficiency, sequencing depth, and pipetting variance. Olink solves this with Normalized Protein eXpression (NPX), an arbitrary log2 unit computed per-analyte using built-in extension and amplification controls plus a bridging normalization sample run on every single plate.
The NPX calculation pipeline applies a sequence of corrections designed to remove every source of technical variance that is not biological:
1. Extension control normalization — a spiked-in, fixed-concentration synthetic oligo pair undergoes the exact same extension and amplification steps as real analytes in every well, correcting for well-to-well pipetting and reaction efficiency differences. 2. Inter-plate control (IPC) normalization — a pooled reference plasma sample is run identically on every plate produced across a study, however large. Each plate's analyte-level offset from the historical IPC median is subtracted, "bridging" plates run months or years apart onto the same scale. 3. Log2 transformation — raw signal (inverted Ct, or normalized read count) is log2-transformed, so that a 1.0 NPX unit difference always represents a 2-fold change in relative protein abundance, regardless of the analyte's baseline signal intensity. 4. LOD flagging — for each analyte, the limit of detection is calculated as 3 standard deviations above the mean of negative (buffer-only) control wells; sample values below LOD are flagged (not deleted) so users can choose their own missingness policy downstream.
The result is an NPX matrix — samples × proteins — where values are directly comparable within an analyte across the entire dataset, but explicitly not comparable in absolute terms between different proteins (NPX is relative, not molar, concentration).
Beyond LOD flagging, every plate and every sample carries automated QC deviation warnings:
• Detection control deviation — flags plates where the extension/detection control signal deviates >0.3 NPX from the expected value, indicating a possible reagent or instrument issue for that whole plate • Incubation control deviation — flags individual samples where the incubation control signal is out of range, often indicating a pipetting error or sample-specific matrix interference (e.g., strong hemolysis) • Assay warning — any single analyte within a sample whose signal falls outside expected bounds
In practice, well-run studies see over 95% of sample-analyte data points pass QC cleanly; the remainder are typically driven by pre-analytical sample handling issues (repeated freeze-thaw, delayed processing, hemolysis) rather than assay failure, which is why standardized biobank collection protocols matter as much as the assay chemistry itself.
The final and highest-value step turns a large NPX matrix into biological and clinical knowledge: which proteins differ between disease and health, which predict future events, and which point to druggable pathways. This requires rigorous statistics, independent replication, and — ideally — orthogonal confirmation with a completely different analytical technology before a candidate is trusted as a real biomarker rather than a batch artifact.
A typical Olink biomarker discovery analysis proceeds through several stages:
• Pre-processing — NPX values below LOD in a large fraction of samples are either excluded per-analyte or imputed; samples failing plate/sample QC are removed entirely • Covariate adjustment — linear or linear mixed-effects models regress NPX on disease status while adjusting for age, sex, BMI, sample storage time, and often plate/batch as a random effect • Multiple-testing correction — with 1,500–5,400 simultaneous tests, Benjamini-Hochberg FDR control (or Bonferroni for very conservative studies) is mandatory; an uncorrected p<0.05 threshold at this scale would produce hundreds of false positives by chance alone • Effect size ranking — candidate biomarkers are ranked by both statistical significance (−log10 p) and effect size (log2 fold-change / NPX difference), commonly visualized as a volcano plot • Machine learning panels — for diagnostic or prognostic use, multi-protein signatures (LASSO, gradient-boosted trees) frequently outperform any single analyte, trading interpretability for discriminative power (AUC)
Discovery cohorts are only the first step: any signal must be tested in at least one independent replication cohort before publication-grade confidence is warranted.
The UK Biobank Pharma Proteomics Project (UKB-PPP), published in Nature (2023), is the largest population-scale application of Olink technology to date: ~2,923 proteins (Explore 3072 panel) were measured in plasma from 54,219 UK Biobank participants, generating one of the largest human plasma proteomic resources ever assembled. The study mapped roughly 14,000 protein quantitative trait loci (pQTLs) — genetic variants associated with circulating protein levels — providing a genome-wide map connecting DNA variation to the plasma proteome and nominating causal drug targets via Mendelian randomization.
Before any single-protein or multi-protein signature is considered a validated biomarker, best practice requires: 1. Independent cohort replication — the association must reproduce in a demographically distinct population 2. Orthogonal technology confirmation — mass spectrometry (targeted MRM/PRM assays) or a conventional immunoassay (ELISA, Luminex) measuring the same protein by a chemically unrelated method should agree directionally with the PEA signal 3. Biological plausibility — the candidate should fit a coherent pathway model, ideally supported by pQTL/Mendelian randomization evidence that the protein is causally upstream of the phenotype, not merely a downstream consequence
The UK Biobank Pharma Proteomics Project profiled 2,923 circulating proteins across 54,219 participants using Olink Explore panels — mapping ~14,000 pQTLs and demonstrating that population-scale, multiplexed PEA proteomics can nominate and genetically triage drug targets at a speed and scale no single-analyte immunoassay could match. Roughly 70–80% of discovery-cohort hits have replicated in independent validation cohorts across published Olink biomarker studies, underscoring both the power and the residual need for orthogonal confirmation before clinical translation.