TMB Incubation Time in ELISA: Why It Makes or Breaks Results

A researcher ran a high-sensitivity porcine insulin sandwich ELISA on EDTA plasma and saw main-run duplicate CVs as high as 29.6% (kit spec: intra-assay CV <8%) together with a standard curve whose lowest point sat only about 0.0018 OD above the blank. Despite an R² >0.99, the curve was compressed because colour development was stopped too early (~6 minutes). Extending TMB development into the validated 10–20 minute window (~15 min, top standard reaching ≥1.5 OD), validating the curve before committing precious samples, and diluting the plasma into the middle of the curve (1:5–1:10) restored interpretable, reproducible results.

A customer measured insulin in porcine EDTA plasma using a high-sensitivity sandwich ELISA. On the main run, five of eight in-range duplicate pairs exceeded the 10% CV acceptance threshold, with the worst pair reaching 29.6% — well beyond the kit’s stated intra-assay CV of <8%. At the same time, the calibration looked superficially fine (R² >0.99) yet the lowest standard barely rose above the blank. The researcher, quite reasonably, asked whether any of the numbers could still be trusted, and whether a hook effect or a bad reagent lot was to blame. This is a textbook case where two symptoms — poor precision and a “flat” low end — share a single upstream cause.

Two views tell the story: the calibration itself, and how the plasma behaves on dilution.

Standard curve (main run) — observed ODs

OD values are the mean of duplicate wells from the main run, rounded; sample labels are anonymised. The top standard (500 pg/mL) reached only ~0.66 blank-corrected OD — the kit’s QC reference at 500 pg/mL is ~2.3, and the internal minimum for the highest standard is ≥1.5 — while the lowest standard sat just 0.0018 above the blank.

Dilution linearity — dilution-corrected insulin (pg/mL)

Values are the customer’s own range-finding readings, rounded to whole pg/mL; sample labels are anonymised. Compared with the undiluted sample the corrected value jumps 2–4.5× — but between 1:5 and 1:10 the samples do dilute linearly (the corrected concentrations hold to ~0.9–1.3×, roughly within ±20%). That 1:5–1:10 band is the reliable window to read in; the big overall drift comes from the neat and low-dilution points on a compressed curve.

In a sandwich ELISA the HRP–TMB reaction is kinetic: horseradish peroxidase turns over TMB substrate continuously, so absorbance climbs with time. Each well’s final OD is therefore a product of two things — how much bound analyte captured detection antibody (what you want to measure) and how long you let the colour develop (what you control). Stop the reaction early and every well is scaled down together, but the effect is not uniform in usefulness: the top standards still give a readable signal, while the lowest standards are pushed down toward the blank, where they can no longer be told apart from background.

This is why the standard curve is the ruler of the whole assay. Its job is to span a usable dynamic range — a top anchor comfortably high (our internal target for a sandwich ELISA is an OD of at least ~1.5, ideally ~2.0–2.5 for the highest standard) and a bottom anchor that is still meaningfully above the blank. A high R² does not guarantee this. R² only measures how well points fit a model; a compressed, near-linear curve can fit beautifully and still be useless, because it has almost no signal-to-blank separation at the low end. When the lowest standard sits ~0.0018 OD above the blank, it is inside the noise, and any concentration read against that region is essentially unquantifiable.

The precision problem follows from the same physics. %CV is the standard deviation of replicates divided by their mean. Near the blank, the mean signal is tiny while the absolute pipetting/optical noise is roughly constant, so that noise becomes a large percentage — CVs balloon exactly where the curve is weakest. An under-developed curve therefore manufactures both symptoms at once: no low-end sensitivity and inflated %CV.

The last principle is dilution linearity (parallelism), and it deserves a closer look because it is where this case gets subtle. Plasma is a complex matrix, and a correctly behaving sample obeys a simple rule: once you multiply back by the dilution factor, every dilution should return the same concentration, agreeing within roughly ±20%. That is the definition of a sample that is diluting linearly and reading on a trustworthy curve. When the dilution-corrected result instead drifts in one direction as you dilute, the assay is telling you that either the sample matrix or the curve — or both — is distorting the numbers.

Two mechanisms push the corrected value up with dilution, which is the direction seen here. The first is low-dilution signal suppression (a matrix effect): concentrated plasma contains binding proteins and other components that mask analyte or interfere with antibody binding, so the neat sample under-reads; diluting the matrix relieves the interference and reveals more of the true signal. The second is the shape of an under-developed curve: undiluted samples give the highest ODs and therefore sit on the flat, poorly-resolved top of a compressed curve, where a small OD change maps to a large concentration change and quantification is least reliable. Diluting moves the sample down into the well-resolved middle of the curve, which is exactly where it should be read. (The mirror-image failure — the high-dose hook effect, where extreme antigen excess depresses signal — would show the opposite pattern, and is not what these data show.)

Mapping the theory onto the numbers, the picture is consistent. The colour reaction was terminated at roughly six minutes, before the standards had reached their working signal. That left the top standard at only ~0.66 blank-corrected OD — well under a third of the QC reference of ~2.3 at 500 pg/mL, and below the internal minimum of 1.5 — and collapsed the whole curve downward, leaving the 7.813 pg/mL point (just 0.0018 above the blank) inside the blank’s noise. With the ruler compressed like that, the low-end concentrations and the calculated %CVs could not be evaluated fairly — the reference itself was not yet valid.

We were careful, though, not to blame everything on development time. Some duplicate disagreement appeared at higher raw ODs too (roughly 19–22% CV on pairs well away from the blank), and that is not explained by under-development. Signal that is high but still irreproducible points to pipetting or mixing variability — inconsistent volumes, bubbles, incomplete mixing, or plate edge effects. That component needs to be assessed on its own, independent of the curve.

The dilution linearity deserves the same care, because it is the part of the data most easily over-interpreted. Across the four representative samples, the dilution-corrected concentration climbed steadily from the neat sample to the 1:10 point — roughly 2.0×, 4.0×, 4.5× and 4.4× higher at 1:10 than undiluted. Read on its own, that upward drift looks like a textbook matrix effect relieved by dilution. But two things make it impossible to draw that conclusion cleanly from this particular run. First, the range-finding standard curve was itself under-developed — its top standards reached only ~0.74–0.81 blank-corrected OD, far short of the working target — so the ruler used to back-calculate every one of these numbers was compressed, and the undiluted samples sat on its least reliable, flattest region. Second, the researcher transparently disclosed that this range-finding series carried its own confounders: a short delay in adding detection antibody to the 1:2 and 1:10 wells, and the use of different stored sample aliquots. Those wells are therefore not strictly comparable to the undiluted and 1:5 wells.

So the honest reading is layered. The direction of the drift is consistent with low-dilution suppression plus a compressed curve, and it points clearly to the practical fix — dilute into the middle of the curve, where 1:5 to 1:10 placed the ODs. And crucially, the higher-dilution points already behave: between 1:5 and 1:10 the corrected concentrations hold roughly steady (~0.9–1.3×, within about ±20%), so the sample is diluting linearly across that band — which is exactly why 1:5–1:10 is the working window to read in. The large overall spread is driven by the neat and low-dilution points, not by that reliable range. What still cannot be treated as a clean measure of non-linearity is the full 2–4.5-fold undiluted-to-1:10 range, until the curve is developed and the disclosed range-finding artefacts are removed. In other words, linearity here is a question to re-open on a valid curve, not a conclusion to bank from this run. Supporting that, the glucagon assay run in parallel on the same samples behaved normally, which keeps the focus on this assay’s colour-development and calibration rather than on the samples themselves.

Can an R² >0.99 still be a bad standard curve?

Yes. R² measures goodness of fit, not dynamic range or sensitivity. A compressed, under-developed curve can fit a model almost perfectly and still fail to separate low concentrations from the blank. Judge a curve by its top-standard OD and its signal-to-blank at the bottom, not by R² alone.

Why does stopping TMB early inflate %CV?

Because %CV is noise relative to signal. Near the blank the mean OD is tiny while the absolute optical and pipetting noise is roughly fixed, so that noise becomes a large percentage. Poor separation of adjacent standards then magnifies the variability of back-calculated concentrations.

Should I stop on the clock or on the visual endpoint?

Both, together. Read the visual endpoint within the validated 10–20 minute window rather than stopping before it. If colour is weak at 10 minutes, extend toward 20; the highest standard should reach the target OD before you add stop solution.

My duplicates disagree even at high OD — is that the curve too?

Not necessarily. Under-development explains poor low-end behaviour, but high-signal duplicates that still disagree point to pipetting, mixing, bubbles or edge effects. Assess that component independently once you have a valid curve.

Why does my sample read higher the more I dilute it?

A dilution-corrected value that climbs with dilution usually means the neat sample is being under-read — either the concentrated matrix is suppressing signal (relieved by dilution) or the undiluted sample sits on the flat, poorly-resolved top of a compressed curve. Fix the curve first, then re-test the dilution series with every dilution handled identically; a sample that dilutes linearly should give the same concentration (within ~±20%) at every dilution.

ELISA Support Hub →

Calculating & Analyzing ELISA Data — standard curves, %CV, spike recovery →

101 ELISA Troubleshooting Tips →

ELISA Controls Guide — linearity of dilution & matrix interference →

Sandwich ELISA: Step-by-Step Protocol & Troubleshooting →

Complete ELISA Guide: Principles, Methods & Kits →

Our PhD-level scientific support team reviews your raw data and helps you get an assay back in range.