Four-Parameter Logistic (4PL) ELISA Standard Curve Fitting

An ELISA standard curve is not a straight line. Absorbance (or fluorescence/luminescence) rises steeply through the mid-range of the assay and flattens at both ends — near zero concentration, where there is no more signal to lose, and near saturation, where every binding site is already occupied. Fitting that S-shape with linear or simple polynomial regression systematically distorts concentrations at the top and bottom of the range even when the mid-range fits well. The four-parameter logistic (4PL) model is the standard fix, and ICH Q2(R2) names it explicitly: the 2024 revision folds calibration-curve linearity into a broader “Response (Calibration Model)” category and splits it into linear versus non-linear, with non-linear covering “4- or 5-parameter logistic” fits “named for S-shaped immunoassay and cell-based assay curves.” This guide assumes a completed plate read — see ELISA protocol basics for the assay steps that produce the raw data — and covers the three decisions that actually determine whether a 4PL fit is trustworthy: how to weight the regression, what counts as acceptable back-calculated accuracy, and when a standard point should be dropped rather than forced into the curve.

What the 4PL model fits, and why linear regression is the wrong tool

The 4PL equation is:

y = D + (A − D) / (1 + (x / C)^B)

where A is the response at zero analyte concentration (the bottom asymptote for a standard sandwich ELISA), D is the response at infinite concentration (the top asymptote, where the signal saturates), C is the concentration at the curve’s inflection point — the EC50, the point of maximum sensitivity — and B is the Hill slope, which controls how steeply the curve transitions between the two asymptotes. All four parameters are estimated simultaneously by non-linear least squares; there is no closed-form solution, so the fit is iterative and depends on starting values, which is why fitting software seeds A and D from the observed low/high response and C from the response midpoint before iterating.

A linear or quadratic fit forced onto this data will look acceptable by R² alone — R² rewards fitting the points with the most spread, which in an ELISA are the mid-range standards, and a sigmoidal curve’s mid-range is close to linear anyway. The distortion shows up specifically at the top and bottom standards, exactly where a linear fit has no mechanism to flatten. That is a different failure mode from the linear calibration curves used for protein assays like Bradford, which genuinely are linear across their working range and don’t need this correction — see building an analytical calibration curve and the Bradford protein assay standard curve for the linear case. R² is consequently a poor diagnostic for a 4PL fit’s quality on its own; back-calculated accuracy (below) is the test that actually catches asymptote distortion.

Choosing a weighting scheme

Unweighted (ordinary) least squares assumes the variance of the response is constant across the whole curve. ELISA data does not behave that way: absorbance readings are heteroscedastic, and the variance of replicate readings typically grows with the signal itself — the high standards, with their large absolute variance, dominate an unweighted fit and pull the curve away from the low standards, which is exactly where an assay’s sensitivity and lower limit of quantitation live. Weighting corrects this by down-weighting points in proportion to their variance so the fit is not dominated by the noisiest end of the range.

The weighting schemes in common use, from mildest to most aggressive correction:

• 1/y — a moderate correction, weighting inversely to the response itself.
• 1/y² — a stronger correction, appropriate when variance grows roughly with the square of the response (constant %CV across the range, which is common for plate-based immunoassays).
• 1/x and 1/x² — weight by nominal concentration instead of observed response; used less often for ELISA specifically since concentration, not response, is the axis with less direct control over variance.

There is no universal correct choice — it depends on the actual variance structure of a given assay/plate reader/detection chemistry combination, which is why fitting software (SoftMax Pro on a plate reader, GraphPad Prism, the drc package in R, or a custom scipy.optimize.curve_fit fit in Python) exposes weighting as a selectable option rather than hard-coding one. The practical way to choose is empirical, not theoretical: fit the curve unweighted first and look at the residuals (fitted minus observed) as a function of concentration. If residuals fan out — wider scatter at high concentration than low — that is the heteroscedasticity signature, and 1/y or 1/y² weighting is warranted. Refit with each candidate weighting and compare back-calculated accuracy (next section) across the whole range, not just the R² of the fit itself; the right weighting is the one that produces the flattest, most uniform back-calculated accuracy from top to bottom of the curve, not the highest R².

Back-calculated accuracy: the test that actually validates the curve

Once a curve is fit, the standard way to check whether that fit is trustworthy is to run every standard’s own observed response back through the fitted equation (solving the 4PL equation for x instead of y) and compare the back-calculated concentration to the standard’s known nominal concentration. This is the direct test of whether the curve actually recovers the values used to build it — a high R² can coexist with a curve that systematically over- or under-estimates concentration at one end of the range, and back-calculation is what catches that.

Report back-calculated accuracy as percent of nominal for every standard: (back-calculated concentration / nominal concentration) × 100 — the same recovery logic used to check any calibration curve, including the linear case covered in LOD vs LOQ calculation. Bioanalytical method validation practice commonly applies a looser accuracy tolerance to ligand-binding assays like ELISA than to chromatographic (LC-MS/GC) methods, reflecting the larger inherent biological and reagent-lot variability of an immunoassay; a widely used convention is accepting standards within roughly ±20% of nominal across the curve, with a wider allowance (commonly ±25%) at the lowest non-zero standard, since variability is intrinsically highest near the assay’s detection floor. Treat the exact percentage as a starting convention to calibrate against your own lab’s or CRO’s validated SOP and any applicable regulatory submission requirements, not a number to cite verbatim without checking the guidance document that actually governs your assay’s intended use — the specific figures above are reported practice, not independently re-verified against a specific FDA/EMA guidance section in this piece.

A curve that fails back-calculated accuracy at every standard, not just the extremes, usually means the weighting is wrong, not that individual points are bad — refit with a different weighting scheme before touching any individual standard. A curve that fails accuracy at one or two standards specifically, especially the top or bottom of the range, while the rest of the curve recovers cleanly, is the pattern that points to dropping an anchor point instead.

When to drop an anchor point

The top and bottom standards on an ELISA curve (the “anchor points”) do disproportionate work: they set the A and D asymptotes that the whole sigmoidal shape is built from, so an unreliable anchor point distorts more than just its own end of the curve — it can pull the inflection point and slope estimate too. That makes an anchor point both the most likely standard to be pipetting- or plate-edge-affected and the most consequential one to get wrong.

The decision to drop an anchor point should be made on evidence, not on convenience because it improves the fit statistics:

• Genuine detector/reagent saturation at the top standard — if the highest standard and the second-highest standard produce nearly identical absorbance (the curve has visibly plateaued before reaching the top standard), that top point is past the assay’s real dynamic range and is legitimately excludable; refit without it and confirm the D asymptote and back-calculated accuracy improve across the rest of the curve.
• A bottom standard indistinguishable from the blank — if the lowest non-zero standard’s response is within the assay’s normal blank-well variability, it is not contributing real information to the A asymptote and can be dropped or, better, re-run at a validated concentration further from the blank.
• An isolated back-calculation failure with no plausible mechanistic cause — a single standard that fails accuracy while its neighbors and the rest of the curve recover cleanly is more often a pipetting, mixing, or transcription error on that one well than a real feature of the assay’s response range; before dropping it, check the raw plate data for that well (edge-well position, a duplicate/triplicate CV outlier) rather than dropping on the fit alone.

Two things this is not a license to do. First, don’t drop a point purely because removing it raises R² or tightens back-calculated accuracy on the remaining points — that is curve-fitting to the outcome you want, not validating the assay, and it hides a real problem (assay range, reagent lot, plate handling) rather than fixing it. Second, don’t drop points until too few remain: a 4PL fit has four parameters to estimate, so a curve needs enough concentrations spanning the range to constrain all four reliably — ICH Q2(R2) sets a minimum of five concentrations for the linear calibration case, and a non-linear 4PL fit with its extra two parameters needs at least that many, in practice more, spread across the full response range including points that genuinely define both asymptotes. If dropping a point brings the curve below what’s needed to constrain A, B, C, and D independently, the fix is re-running the plate with a corrected standard, not shipping a curve fit to too few points.

Frequently asked questions

Can I just use a 5PL fit instead of 4PL?

A five-parameter logistic (5PL) adds an asymmetry parameter, letting the curve approach its two asymptotes at different rates rather than assuming a symmetric sigmoid around the inflection point. ICH Q2(R2) groups 4PL and 5PL together as the same category of non-linear immunoassay fit. Whether 5PL is worth the extra parameter depends on whether your specific assay’s curve is visibly asymmetric on a semi-log plot; adding a parameter to a curve that doesn’t need it just makes the fit less stable with the same number of standards, so don’t switch to 5PL by default — check for real asymmetry first.

What R² should an ELISA standard curve hit?

R² is a weak diagnostic for a 4PL fit specifically, for the reason covered above — it’s dominated by mid-range points and can stay high even while the curve misfits the asymptotes. Treat a high R² as necessary, not sufficient; back-calculated accuracy across every standard, not just the correlation coefficient, is the test that actually validates the fit.

Why does my curve fail to converge?

Non-linear least squares iterates from starting estimates for A, B, C, and D; poor starting values, too few standards to constrain four parameters, or a genuinely flat/non-sigmoidal dataset (no real top or bottom plateau in the data) are the most common causes. Confirm the plate actually shows a plateau at both ends before troubleshooting the fitting software itself.

Does weighting change the EC50 (the C parameter)?

Yes, potentially — weighting changes which points dominate the fit, and the inflection point is influenced by all four parameters jointly, so a different weighting scheme can shift the estimated C along with A and D. This is one more reason to choose weighting from the residual pattern and back-calculated accuracy across the whole curve, not from whichever weighting happens to produce the C value expected going in.