ELISA for Peptide Quantification: Cross-Reactivity and Matrix Effects

An ELISA does not measure a peptide. It measures how much signal a pair of antibodies — or one antibody competing against a labeled tracer — produces from a well, and then converts that signal into a concentration using a curve built from a calibrator you chose, diluted in a fluid you chose. Every link in that chain can hold for the calibrator and break for the sample. What comes out is a measure of immunoreactivity under one set of reagents, and the distance between immunoreactivity and molar concentration is the subject of this guide.

The gap is wider for peptides than for most analytes. Peptides are small, so there is less surface on which to land two antibodies. They are generated by cleaving larger precursors and cleared by cleaving further, so a sample commonly holds a family of species differing from the intended analyte by a few residues at one terminus. And they are often carried by binding proteins that cover the very regions an antibody was raised against.

The two formats fail in opposite directions

A sandwich ELISA captures the analyte on an immobilized antibody and detects it with a second, labeled antibody against a different epitope. Signal rises with concentration. Because two binding events are required, a species carrying only one of the two epitopes should not be counted — which is why the sandwich is usually described as the more specific format. A competitive ELISA uses one antibody and makes the sample compete against a labeled or immobilized version of the analyte. Signal falls as concentration rises, and because only one binding event is required, anything occupying that single site is counted.

The practical consequence is that a sandwich requires two epitopes to be present, non-overlapping, and simultaneously accessible on the same molecule. Many research peptides are too short to satisfy that, which is why competitive formats remain the default at the small end of the size range. This is a structural constraint rather than a finding from any single study, and we state it as convention.

Requiring two epitopes buys specificity, but it does not guarantee it. Three commercial glucagon sandwich ELISAs were compared head to head: the first showed 22% cross-reactivity with glicentin along with insufficient sensitivity, while the second and third showed minor cross-reactivity of 1–5% with oxyntomodulin and glicentin, and only one of the three was judged by those authors to be suitable for the plasma samples in that study ^[1]. A later group developing a new glucagon sandwich ELISA reported cross-reactivities below 1% and a lower measuring range of 0.31 pmol/L, explicitly motivated by cross-reactivity of an existing sandwich ELISA with proglucagon-derived peptides ^[2]. Two antibodies raised against two ends of a peptide will both bind a precursor that still contains both ends.

Cross-reactivity is a ratio, not a property

A datasheet figure such as "<1% cross-reactivity with analog X" is half a statement. Interference is cross-reactivity multiplied by the molar excess of the interferent, and datasheets rarely tell you what that excess will be in your matrix.

Take the 22% figure reported above ^[1] and suppose, hypothetically, that the cross-reactant sits at 5 nmol/L while the analyte sits at 1 nmol/L. The apparent concentration is 1 + (0.22 × 5) = 1 + 1.1 = 2.1 nmol/L. The error relative to the true value — and the true value is the denominator — is (2.1 − 1.0) / 1.0 = 110% high. Run the same arithmetic at the 1–5% cross-reactivity also reported in that comparison ^[1]: at 5%, 1 + (0.05 × 5) = 1.25 nmol/L, or 25% high; at 1%, 1 + (0.01 × 5) = 1.05 nmol/L, or 5% high.

Now hold cross-reactivity at 1% and raise the molar excess to 100-fold: 1 + (0.01 × 100) = 2.0 nmol/L, which is 100% high. A 1% cross-reactant present at a hundredfold excess doubles the answer. This is why the interesting question about a peptide assay is not "how specific is the antibody" but "what else is in the tube, and at what ratio". A review of immunoassays for the incretin peptides makes the same point from the analyte side: processing of the precursor molecules gives rise to a number of different peptides that cross-react with antisera raised against the two hormones, and the reviewers state that the degradation products are the major circulating forms, accounting for 60–80% of total immunoreactive GLP-1 and GIP in peripheral plasma, while intact forms can be very low and in some cases barely detectable ^[3]. When the fragment is the dominant immunoreactive species, a percentage on a datasheet is not a reassurance.

Fragments can also make a result too low

The opposite failure mode is the one that surprises people. Fragments that bind the antibody but cannot complete the sandwich, or that are recognized less efficiently than the calibrator, suppress signal rather than add to it.

The clearest demonstration comes from outside peptide work. Using LC-MS/MS to profile hydrolyzed gluten in beer, one group found that certain hordein classes were prone to hydrolysis, and that beers containing high levels of B-hordein fragments gave near-zero values by ELISA; the persisting hydrolyzed fragments showed a dose-dependent suppression of the ELISA measurement despite the assay being calibrated with a hordein standard ^[4]. The calibrator was intact, the sample was not, and the assay reported the difference as absence rather than as a change in form.

The same logic applies to any sandwich assay for a peptide that is present partly as truncated forms. A species that occupies the capture antibody without presenting the detection epitope acts as a competitive inhibitor inside a format chosen for its specificity.

Epitope availability: the antibody sees a surface, not a sequence

An antibody binds a conformation in a context. Anything covering, distorting or removing that surface reduces the count, and the calibrator — clean peptide in clean buffer — carries none of those obstructions.

Carrier proteins are the standard example. In a review of clinical assays for IGF-1, the authors note that the great majority of circulating IGF-1 is bound in a ternary complex with IGF binding protein 3 and the acid labile subunit, that binding proteins are a known source of interference in immunoassays, and that assay methodology therefore has to dissociate IGF-1 from the binding proteins before quantitation; they describe immunoassays as relying on antibody specificity toward IGF-1 and on differential cross-reactivity toward the binding proteins ^[5]. A dissociation step that clears a calibrator spiked into buffer may not clear material that has equilibrated with a real matrix — an expectation from the chemistry rather than a result we can cite.

Epitope selection also works in the other direction, as a deliberate tool. A sandwich ELISA for human tissue prokallikrein was built with monospecific anti-peptide antibodies raised against propeptide derivatives — that is, against a region present in the precursor and absent from the mature enzyme. The authors reported a minimum detectable concentration of 60 pg/mL, no cross-reactivity or interference with mature tissue kallikrein, intra- and inter-assay precision varying from 8% to 15%, mean recovery of added zymogen of 104%, and dilution curves showing good linearity in urine samples ^[6]. Discrimination between two closely related species came from choosing the epitope, not from the format alone.

What the sample brings with it

Matrix interference is anything in the sample that changes the signal without changing the analyte. Three categories recur:

• Bridging antibodies. Antibodies in the sample that recognize immunoglobulins from the species the assay reagents came from can link capture to detection with no analyte present. In a B-type natriuretic peptide sandwich ELISA built from a mouse monoclonal and a rabbit polyclonal against synthetic peptides, heterophilic antibodies in dog plasma were shown to be capable of causing false-positive results; those antibodies reacted with blood proteins from several species at relatively low titers and comprised both IgG and IgM, and a combination of protein A agarose precipitation with mouse IgG blocking antibody was effective in eliminating the false positives while retaining adequate test performance ^[7]. This artifact belongs specifically to the sandwich format — it manufactures a bridge, and a competitive assay has no bridge to manufacture.
• Reagent chemistry. Assays built on biotin–streptavidin capture are vulnerable to free biotin in the specimen competing for streptavidin sites. A review of that chemistry notes that elevated biotin levels vary between specimens and that the resulting interference cannot easily be predicted or quantified, and that manufacturers investigate biotin interference against a threshold of 3510 ng/mL stipulated by the FDA ^[8]. The relevant question for a bench assay is whether the sample preparation could introduce free biotin at all.
• Bulk physicochemical differences. Protein content, lipid, ionic strength, anticoagulant and pH all shift binding kinetics and non-specific adsorption. These effects are detected rather than predicted, which is what the next section is about.

Parallelism is the experiment that finds the problem

Parallelism asks whether serial dilutions of a real sample, once corrected for the dilution factor, back-calculate to the same concentration. A review of parallelism in ligand-binding assay development describes it as an essential experiment characterizing relative accuracy, and notes that by assessing the effects of dilution on quantitation of endogenous analyte in matrix, a single experiment addresses selectivity, matrix effects, the minimum required dilution, endogenous levels and the lower limit of quantification ^[9].

Work the arithmetic. Suppose a neat sample reads 200 pg/mL after dilution correction, a 1:2 dilution back-calculates to 170 pg/mL, 1:4 to 150 pg/mL, and 1:8 to 140 pg/mL. Indexed to the neat result as the denominator, recoveries are 170/200 = 85%, 150/200 = 75% and 140/200 = 70% — a monotonic drift, not scatter. Index the same four numbers to the most-diluted point instead and the neat sample reads 200/140 = 143% of it. The data have not changed; the choice of denominator has. Since the neat result is the one under suspicion, indexing to it is the framing that makes the drift visible.

Nonparallelism tells you something is wrong; it does not tell you what. It can be matrix suppression, a cross-reactant diluting out at a different rate than the analyte, or a format artifact. One group profiling a therapeutic monoclonal antibody against a soluble target in cynomolgus monkeys found a substantial discrepancy between a generic anti-human IgG capture and detection assay and a target bridging assay, with the target bridging assay giving dose- and time-dependent underquantification and a lack of parallelism; concentrations from a target capture assay and an anti-idiotypic bridging assay agreed closely with the generic assay and showed parallelism with high precision across several dilutions ^[10]. Those authors present this as a practical attempt to overcome nonparallelism by changing format and reagents rather than the sample.

Designing the standard curve so it answers the right question

The calibrator defines what the number means, and the two most consequential decisions are what the calibrator is and what it is dissolved in.

Calibrator matrix

Calibrators in assay buffer and samples in a biological matrix are not the same measurement. The GLP-1 kit comparison shows how large the gap can be: synthetic GLP-1 isoforms at known concentrations were added both to the buffer supplied with each of ten kits and to human plasma, and recoveries determined in each. One total-GLP-1 kit detected all six isoforms but with incomplete recovery of non-active forms, especially in plasma; another detected all isoforms in buffer but mainly amidated forms in plasma; a third showed incomplete recovery in plasma ^[11]. The authors concluded that specificity and sensitivity vary considerably between commercially available kits and that this should be taken into account when selecting an assay and when comparing data across studies ^[11]. Notably, the response pattern across the clinical sample series was similar between the kits tested while the absolute concentrations measured varied ^[11] — a reminder that an assay can rank samples correctly and still be wrong about their values.

Minimum required dilution

If a matrix must be diluted before it behaves, that dilution multiplies the working sensitivity. An assay with an LLOQ of 10 pg/mL in the well, run at a minimum required dilution of 1:4, cannot quantify below 10 × 4 = 40 pg/mL in the original sample, and a well reading of 30 pg/mL corresponds to 30 × 4 = 120 pg/mL neat. The dilution that fixes the interference is the same dilution that costs you the bottom of the range; this trade-off is one of the parameters the parallelism experiment is used to set ^[9].

Curve shape and where the error lives

Immunoassay response is sigmoidal, and four- or five-parameter logistic fits are standard; back-calculation should stay inside the calibrated interval, because extrapolation past the top or bottom standard assumes a curve shape that has not been demonstrated. We state that as convention rather than as a cited finding. One curve-shape artifact does have direct support. Authors developing a kinetics-based mitigation state that sandwich-type immunoassays are often limited by the high-dose hook effect, which causes falsely low results when analytes are present at very high concentrations; their demonstration was in a lateral flow sandwich format, and the generality they claim for their technique is to hook-effect-limited sandwich lateral flow assays, not to plate ELISAs ^[12]. A single well cannot distinguish a low reading on the ascending limb from a low reading past the hook, which is why a low result on a sample expected to be concentrated should be re-run diluted.

What is not known, and what to do about it

For most research peptides there is no reference material with an assigned value, no interlaboratory comparison, and no published cross-reactivity panel covering the specific analogs and truncation products likely to be present. Manufacturer cross-reactivity tables are generally generated against a short list of related compounds in buffer — the condition least likely to reveal matrix-dependent behavior.

That leaves a short list worth doing before believing a figure, most of it bench convention rather than published method: run a dilution series on every new matrix and inspect it for drift, not just for an r² value; spike a known amount of peptide into the actual matrix and recover it; run a matrix blank; and where the stakes justify it, confirm against a method that separates by mass rather than by binding, as several of the studies cited here did ^{[2, 4]}. An ELISA result that agrees with an orthogonal method across a dilution series is a different kind of claim from one that has only ever been compared to its own calibrator.