A standard curve fitted with an excellent R² tells you the assay works in buffer. It says nothing about whether the same antibody pair reads correctly in serum, plasma, cell lysate, urine or sediment extract. Two experiments answer that question: spike recovery and dilution linearity. If your ELISA data has ever produced concentrations that changed when you diluted the sample, this is the reason.

Why matrix matters more than the curve

Sample matrices interfere in ways that never appear in a calibrator prepared in assay buffer:

Binding proteins and soluble receptors sequester the analyte, so only a fraction is available to the capture antibody.
Heterophilic antibodies and rheumatoid factor bridge capture and detection antibodies without any analyte, producing a false positive signal.
High total protein, lipids or salts shift the local optical background and alter binding kinetics.
Autoantibodies against the analyte mask epitopes.

None of these change the calibration curve. They change the relationship between the sample and the curve — which is exactly what the two experiments below quantify.

Experiment 1 — spike recovery

Purpose: does a known amount of analyte, added into the real matrix, read back correctly?

Protocol:

Pick at least three independent matrix samples (different donors/animals/sites), ideally with low endogenous analyte.
Prepare a spike stock at high concentration so the spike volume stays ≤ 10 % of the total — otherwise you dilute the matrix and defeat the purpose.
Spike at three levels covering low, mid and high working range (typically ~3×LOQ, mid-curve, ~80 % of ULOQ).
Run the unspiked sample in the same plate and the same dilution.

Calculation:

Recovery (%) = 100 × (C_spiked − C_unspiked) / C_spike_added

Level	Endogenous	Expected added	Measured spiked	Recovery
Low	12 pg/mL	50 pg/mL	58 pg/mL	92 %
Mid	12 pg/mL	250 pg/mL	231 pg/mL	87.6 %
High	12 pg/mL	800 pg/mL	690 pg/mL	84.8 %

Acceptance: 80–120 % per level for most immunoassays (85–115 % for regulated ligand-binding assays; see the FDA/EMA bioanalytical method validation guidance). Mean recovery near 100 % with wide scatter is not a pass — evaluate each level separately.

A systematic bias in one direction across all levels indicates a proportional matrix effect and often a curve-slope problem. Recovery that fails only at the low level usually means the LOQ is optimistic in that matrix, not in buffer.

Experiment 2 — dilution linearity and the minimum required dilution

Purpose: does a sample give the same back-calculated concentration at every dilution?

Protocol: take a high sample (endogenous or spiked), prepare a serial 2-fold dilution series in assay diluent spanning at least four points inside the calibration range, and back-calculate each point after multiplying by its dilution factor.

Recovery of dilution (%) = 100 × (C_measured × DF) / C_reference

Dilution	Raw result	× DF	Recovery vs 1:4
1:2	940 pg/mL	1880	128 %
1:4	367 pg/mL	1468	100 % (ref)
1:8	180 pg/mL	1440	98 %
1:16	92 pg/mL	1472	100 %

Interpretation of that table: 1:2 is inside a matrix-interference zone; from 1:4 onwards the results are dilution-independent. The minimum required dilution (MRD) is therefore 1:4, and every sample in the study must be run at 1:4 or higher — no exceptions, even for low samples.

Acceptance: each dilution within 80–120 % (±15 % for regulated work) of the reference dilution, and CV of the dilution-corrected values ≤ 20 %.

Non-linearity patterns and what causes them

Values fall with less dilution (concentration increases on dilution): interference by binding proteins or an inhibitory matrix component. Increase MRD.
Values rise with less dilution: non-specific signal added by the matrix, or an antibody cross-reactant.
A curve that turns over at high concentration: the hook effect (prozone) in one-step sandwich formats. High analyte saturates both antibodies separately, so the signal falls again — a very high sample can read as mid-range. The only defence is to run at least two dilutions of every suspicious sample, or to switch to a two-step wash protocol.

Parallelism — the version for real samples

Dilution linearity uses a spiked sample; parallelism repeats the same logic with endogenous high samples and compares the shape of the sample dilution curve with the calibrator curve. If the two curves are not parallel, the sample and the calibrator do not behave as the same substance, and a single-point back-calculation is not defensible. Formal parallelism testing (common slope, F-test on the interaction term) is standard practice in relative-potency work under Ph. Eur. 5.3, and it is the correct tool when you need statistics rather than a 80–120 % rule of thumb.

Reporting checklist

State all of the following, or the validation is not reproducible:

Matrix type, number of independent lots and how they were screened.
Spike levels with nominal concentrations, and recovery per level and per lot.
Diluent composition used for the dilution series.
Dilution factors tested and the resulting MRD, with the criterion used.
Whether hook-effect testing was performed and up to which concentration.
Curve model (4PL/5PL), weighting, and the calibration range in which all of the above holds.

Doing this in Labreador

Labreador handles the analytical part of these workflows in the browser, with no upload:

The ELISA module fits 4PL/5PL curves, back-calculates standards and samples, applies per-well dilution factors, and flags results outside the calibrated range instead of silently extrapolating.
Method Health checks report recovery and bias for back-calculated standards, low dynamic range (ΔOD between top and bottom standard), and unstable confidence intervals — the same failure signatures that matrix problems produce.
The PLA module performs formal parallelism and relative-potency analysis according to Ph. Eur. 5.3 when a rule-of-thumb dilution check is not sufficient.

All processing is local: raw plate data never leaves your device.

References

FDA (2018). Bioanalytical Method Validation — Guidance for Industry.
EMA (2011). Guideline on bioanalytical method validation. EMEA/CHMP/EWP/192217/2009.
Andreasson, U., et al. (2015). A practical guide to immunoassay method validation. Frontiers in Neurology, 6, 179.
European Pharmacopoeia, general chapter 5.3: Statistical analysis of results of biological assays and tests.
Tate, J., & Ward, G. (2004). Interferences in immunoassay. Clinical Biochemist Reviews, 25(2), 105-120.

Cite Labreador

If Labreador supported your analysis, please cite it: Labreador - Bioassay Analysis Platform (Version 1.6.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.21676846

Spike Recovery and Dilution Linearity in ELISA: How to Prove Your Matrix Is Not Lying