Skip to main content
Lab Grimoire
TW EN
Coffee
Lab Methods

How IC50 and EC50 Are Calculated: 4PL Curve Fitting Explained

How IC50 and EC50 come from a 4PL fit: the four parameters, relative vs absolute IC50, a worked 8-dose example, extrapolation, and what to report.

Author
CY
Published
Biomedical Principles and Techniques · Part 4 of 4

An IC50 is read from a four-parameter logistic (4PL) curve fitted to dose and response data by nonlinear least squares. The midpoint that the fit returns is the relative IC50: the dose where the response sits halfway between the two fitted plateaus. The absolute IC50, the dose that gives 50% of the control response, has to be solved from the curve afterwards. Whether either number is fit to report depends on whether the doses cover both plateaus and on how wide the confidence interval is.

The 4PL model and its four parameters

Plotted against log dose, dose–response data usually form an S shape. At low doses the response sits on one plateau, at high doses on the other, with a transition zone between them. The 4PL model describes that curve with four parameters:

Y = Bottom + (Top − Bottom) / (1 + 10^((log10(IC50) − log10(X)) × HillSlope))

  • X: dose or concentration; it must be greater than 0.
  • Y: response, such as viability, percent inhibition, or absorbance.
  • Bottom: the lower plateau, the lowest level the response approaches.
  • Top: the upper plateau, the highest level the response approaches.
  • IC50: the midpoint, the dose at which the response equals (Top + Bottom) / 2.
  • HillSlope: the Hill slope, a unitless measure of how steep the transition zone is.

Fitting means finding the parameter values that minimize the sum of squared differences between the curve and the measured responses. This is nonlinear regression. It has no closed-form solution, so software starts from initial values and iterates until the fit converges. All four parameters are estimated from the data, so you need more than four points; n points leave n − 4 degrees of freedom.

The midpoint is called IC50 for an inhibition curve and EC50 for a curve whose signal rises with dose. Both come from the same equation and the same parameter.

Size and sign of the Hill slope

When the Hill slope has an absolute value of 1, the response takes an 81-fold change in dose to move from 10% to 90% of the transition. A larger absolute value means a steeper transition, with the same change packed into a narrower dose range.

With the equation as written above, the Hill slope is negative for an inhibition curve and positive for a rising curve. Some software writes the exponent as (log10(X) − log10(IC50)), which flips the sign. Compare absolute values across programs, and state the equation in your report. The Hill slope describes the shape of the curve; the number alone does not tell you how many binding sites are involved.

Why dose goes on a log axis

  1. Doses are usually prepared by serial dilution, so neighbouring doses differ by a constant factor and are evenly spaced on a log axis.
  2. The 4PL curve is symmetric around its midpoint on a log-dose axis. On a linear axis, the low doses crowd together at the left edge.
  3. The fit estimates log10(IC50). Its uncertainty is closer to symmetric, so the confidence interval is calculated on the log scale and then converted back to the original units, where it is no longer symmetric.

A dose of zero has no logarithm and cannot be placed on the curve. The untreated control defines 100%; it is not entered as a dose point.

A worked example with eight doses

The table below contains illustrative data, not results from a real experiment. The scenario is a cell viability assay. Responses are expressed as a percentage of the vehicle control (vehicle control = 100%, no-cell blank = 0%), with one mean value per dose.

Dose (µM) Measured viability (%) Fitted value (%) Residual
0.01 98.9 99.81 −0.91
0.03 99.6 98.48 1.12
0.1 93.2 93.29 −0.09
0.3 79.5 79.71 −0.21
1 54.1 54.03 0.07
3 35.3 35.11 0.19
10 26.0 26.42 −0.42
30 24.3 24.06 0.24

Fit results

Fitting the equation above with scipy.optimize.curve_fit in Python (unweighted nonlinear least squares) gives:

Parameter Estimate Standard error
Bottom 23.12% 0.73
Top 100.32% 0.64
log10(IC50) −0.1494 0.0158
IC50 0.709 µM 95% CI 0.641–0.785 µM
HillSlope −1.174 0.047

R² is 0.9997, the residual standard error is 0.775 percentage points, and there are 8 − 4 = 4 degrees of freedom.

The confidence interval is log10(IC50) ± t × standard error, where t is the critical value of the t distribution with 4 degrees of freedom (2.776), raised to a power of 10 at the end. A normal approximation (critical value 1.960) gives a narrower interval of 0.660–0.762 µM. The Lab Grimoire dose–response tool uses the normal approximation, which is why a report should state how the interval was calculated.

How far apart the relative and absolute IC50 are

The fitted value of 0.709 µM is the relative IC50. At that dose the curve gives (100.32 + 23.12) / 2 = 61.72%, halfway between the two plateaus. Viability is still about 62% there; it has not fallen to half of the control.

The absolute IC50 is the dose at which the curve equals 50% of the control. Set Y = 50 and solve the equation for X:

X = IC50 × ((Top − Bottom) / (50 − Bottom) − 1)^(−1 / HillSlope)

With the fitted values: 0.709 × 1.872^0.852 = 0.709 × 1.706 ≈ 1.209 µM.

Same data, same curve, and the absolute IC50 is 1.71 times the relative IC50. The gap comes from the lower plateau stopping at 23% and not reaching 0%. The higher the lower plateau, the larger the gap. When the lower plateau is above 50%, the curve never reaches 50% and no absolute IC50 exists. The two numbers coincide only when Top is 100 and Bottom is 0.

Some reports interpolate between the two points on either side of 50%. A straight line on log dose between 1 µM (54.1%) and 3 µM (35.3%) in the example gives 1.271 µM, close to the 1.209 µM solved from the curve. But interpolation uses only two points, its result shifts with the position and measurement error of those two doses, and it has no standard error or confidence interval to report. A 4PL fit uses every data point and gives the uncertainty of each parameter along with the estimate.

Choosing between the two definitions

Sebaugh (2011) gives the following guidelines. An assay with no stable 100% control must use the relative IC50. An assay with a stable 100% control, but with possibly more than 5% error in the estimate of the 50% control mean, should also use the relative IC50. An assay shown to have an accurate, stable 100% control and less than 5% error in that estimate can use the absolute IC50. The 50% control here is the mean of the 0% and 100% controls.

Rescaling the data so the minimum is 0 and the maximum is 100 does not turn a relative IC50 into an absolute one. It only moves the lower plateau close to 0 on paper and uses no information from the controls. The two kinds of IC50 should not be compared in the same table, and a report must say which one it gives.

Dose design and extrapolation

Number and range of doses

Number. Four parameters need at least five points to leave any degrees of freedom. A common design uses 8 to 12 doses in 3-fold or half-log (about 3.16-fold) serial dilutions. The Lab Grimoire tool requires at least 5 doses above zero and shows a notice below 8.

Range. The data should cover the lower plateau, the transition zone, and the upper plateau. The tool shows a notice when the dose range is narrower than 1.5 log10 units. The example spans 3.48 log10 units. Aim to place the expected IC50 near the middle of the dose range so that enough points describe the plateau on each side. For a compound tested for the first time, a wide dilution factor can locate the IC50 roughly before a second run with closer spacing.

Points at both ends. According to Sebaugh (2011), a relative IC50 should be reported only when at least two doses lie beyond each of the lower and upper bend points. An absolute IC50 needs at least two doses with a predicted response below 50% and at least two above 50%. In the example, the fitted values at 3, 10, and 30 µM (35.11%, 26.42%, 24.06%) are below 50%, and the other five doses are above it.

Without a plateau, the midpoint is extrapolated

If the highest dose is still in the transition zone, no measurement supports Bottom; the curve projects it from its shape. The midpoint is defined as halfway between Top and Bottom, so an uncertain Bottom makes the IC50 uncertain too. An IC50 obtained this way is an extrapolated value.

Remove the three highest doses from the example and refit the five points from 0.01 to 1 µM:

Item 8 doses 5 doses
Bottom (standard error) 23.12% (0.73) 32.97% (19.58)
IC50 0.709 µM 0.556 µM
95% CI of IC50 0.641–0.785 µM 0.001–334 µM
R² 0.9997 0.9988

Five points fitted with four parameters leave 1 degree of freedom. The standard error of the lower plateau grows from 0.73 to 19.58, and the confidence interval spans more than five orders of magnitude, while R² barely moves.

The Lab Grimoire tool applies one fixed rule: when a fitted plateau lies outside the range of the measured responses by more than 10% of that range, the result is marked as extrapolated. For the five-point data, the measured range is 54.1% to 99.6%, 10% of that range is 4.55, and the lower limit is 49.55%. The fitted Bottom of 32.97% is below the limit, so the result is marked.

The remedy is to add higher doses until the plateau appears in the data. If solubility or toxicity rules that out, report "IC50 greater than the highest dose tested" and leave the extrapolated number out. Another option is to fix a plateau from the controls, for example Bottom = 0. That is an assumption; it needs a justification and a line in the report. For viability and inhibition data, the tool also fits a three-parameter model with Bottom fixed at 0 and compares it with the 4PL fit by an F test.

Fit quality: R², residuals, and the confidence interval

R² is almost always high. Sigmoid data cover a wide range of responses, so any curve that passes roughly through the points explains most of the variance. The five-point data in the previous section have an R² of 0.9988, yet the IC50 cannot be pinned down. Spiess and Neumeyer (2010) evaluated R² by Monte Carlo simulation and concluded that it is an inadequate measure for nonlinear models. The Lab Grimoire tool shows a notice when R² falls below 0.95; an R² above 0.95 does not mean the IC50 is ready to report.

Check the residuals. A residual is the measured value minus the fitted value. With a suitable model, residuals scatter around zero, their signs show no trend along the dose axis, and their size is similar to the spread between replicate wells. In the example, the largest of the eight residuals is 1.12 percentage points, and the signs do not cluster in any part of the dose range. If the residuals are all positive at low doses and all negative at high doses, or the reverse, the curve shape does not match the data. Three common causes are an asymmetric curve, a response that reverses at high doses, and a plateau fixed at the wrong value.

Check the confidence interval. The confidence interval expresses the uncertainty of the IC50 as numbers. An interval that spans an order of magnitude means the IC50 is only a rough estimate. Look at the standard errors of Bottom and Top as well: when one of them is very large, the data usually do not reach that plateau.

The confidence interval from a single curve reflects only the precision of that one fit. It does not include variation between batches or days. Replicate wells on one plate are technical replicates and do not count as independent experiments. To describe reproducibility between experiments, repeat the experiment independently and pool the log10(IC50) values.

Check that the parameters are plausible. A Top far above the control, a negative Bottom for viability, or a Hill slope with a very large absolute value all indicate that the data cannot determine the parameters.

When to use 5PL

The 4PL model assumes the curve is symmetric around its midpoint on a log-dose axis. Immunoassay standard curves are often asymmetric, with the two bends curving to different degrees. The five-parameter logistic (5PL) model adds an asymmetry parameter; Gottschalk and Dunn (2005) characterized 5PL and compared it with 4PL.

Switch to 5PL when three conditions hold together: the residuals show runs of bias under 4PL, there are enough points to estimate one more parameter, and the purpose is to back-calculate sample concentrations from readings. In 5PL the midpoint parameter no longer equals the half-response dose, so an EC50 has to be solved from the curve separately. The Lab Grimoire tool fits 4PL only.

To back-calculate sample concentrations, including dilution, from a 4PL standard curve, use the ELISA standard-curve tool.

What to report in a paper

  • Model and equation: write out the 4PL equation, including the sign convention for the Hill slope.
  • Definition of IC50: relative or absolute; for an absolute IC50, state what the 0% and 100% controls are.
  • IC50 with its 95% confidence interval: say how the interval was calculated. When pooling independent experiments, average log10(IC50).
  • Other parameters: Top, Bottom, and Hill slope; which parameters were fixed, at what value, and on what basis.
  • Dose design: number of doses, lowest and highest dose, dilution factor, replicates per dose, and number of independent experiments.
  • Data handling: normalization method, weighting, and any excluded points with the reason.
  • Fit quality: residual standard error or a residual plot; R² may be included but cannot stand alone.
  • Extrapolation: flag it clearly when a plateau is missing, or report "greater than the highest dose tested".
  • Software: name, version, and the function used.

Run it on your own data

The 4PL dose–response tool fits the model in the browser with R (webR), using nls and SSfpl. It reports IC50 or EC50, the 95% confidence interval, Hill slope, R², residuals, and the extrapolation flag, and it provides code you can rerun in desktop R. The R package drc described by Ritz and colleagues (2015) is another common choice on the desktop.

When the data have no plateau and the response changes linearly with dose across the tested range, the 4PL parameters cannot be determined, and the linear regression tool is the better fit.

References

  • Sebaugh JL. Guidelines for accurate EC50/IC50 estimation. Pharm Stat. 2011;10(2):128–134. doi:10.1002/pst.426
  • Ritz C, Baty F, Streibig JC, Gerhard D. Dose-Response Analysis Using R. PLoS ONE. 2015;10(12):e0146021. doi:10.1371/journal.pone.0146021
  • Motulsky H, Christopoulos A. Fitting Models to Biological Data Using Linear and Nonlinear Regression: A Practical Guide to Curve Fitting. Oxford University Press; 2004. doi:10.1093/oso/9780195171792.001.0001
  • Campbell RM, Dymshitz J, Eastwood BJ, et al. Data Standardization for Results Management. In: Assay Guidance Manual [Internet]. Bethesda (MD): Eli Lilly & Company and the National Center for Advancing Translational Sciences; 2012 (updated 2017). https://www.ncbi.nlm.nih.gov/books/NBK91993/
  • Gottschalk PG, Dunn JR. The five-parameter logistic: a characterization and comparison with the four-parameter logistic. Anal Biochem. 2005;343(1):54–65. doi:10.1016/j.ab.2005.04.035
  • Spiess AN, Neumeyer N. An evaluation of R2 as an inadequate measure for nonlinear models in pharmacological and biochemical research: a Monte Carlo approach. BMC Pharmacol. 2010;10:6. doi:10.1186/1471-2210-10-6

References

  1. Sebaugh JL. Guidelines for accurate EC50/IC50 estimation. Pharm Stat. 2011;10(2):128-134.
  2. Ritz C, Baty F, Streibig JC, Gerhard D. Dose-Response Analysis Using R. PLoS ONE. 2015;10(12):e0146021.
  3. Motulsky H, Christopoulos A. Fitting Models to Biological Data Using Linear and Nonlinear Regression: A Practical Guide to Curve Fitting. Oxford University Press; 2004.
  4. Campbell RM, Dymshitz J, Eastwood BJ, et al. Data Standardization for Results Management. In: Assay Guidance Manual. Bethesda (MD): Eli Lilly & Company and NCATS; 2012 (updated 2017). https://www.ncbi.nlm.nih.gov/books/NBK91993/
  5. Gottschalk PG, Dunn JR. The five-parameter logistic: a characterization and comparison with the four-parameter logistic. Anal Biochem. 2005;343(1):54-65.
  6. Spiess AN, Neumeyer N. An evaluation of R2 as an inadequate measure for nonlinear models in pharmacological and biochemical research: a Monte Carlo approach. BMC Pharmacol. 2010;10:6.

Frequently Asked Questions

How is IC50 calculated?

Dose and response data are fitted to a 4PL curve by nonlinear least squares. The midpoint parameter of the curve is the IC50: the dose where the response is halfway between the two plateaus. The fit works on log dose, and the confidence interval is calculated on the log scale and converted back.

What is the difference between relative and absolute IC50?

The relative IC50 is the dose where the response is halfway between the fitted plateaus; the 4PL fit returns it directly. The absolute IC50 is the dose that gives 50% of the control response and is solved from the curve. They differ when the lower plateau is not 0: in the worked example the relative IC50 is 0.709 µM and the absolute IC50 is 1.209 µM.

How many doses does a 4PL fit need?

Four parameters need at least five points. A common design uses 8 to 12 doses that cover the lower plateau, the transition zone, and the upper plateau. Sebaugh (2011) advises reporting a relative IC50 only when at least two doses lie beyond each bend point.

Does a high R² mean the IC50 is reliable?

No. R² is almost always high for sigmoid data. In the worked example, removing the three highest doses leaves R² at 0.9988 while the 95% confidence interval of the IC50 runs from 0.001 to 334 µM. Check the residuals, the confidence interval, and whether the plateaus appear in the data.

Found this useful? RSS · X · Buy me a coffee