A complete concentration-response workflow

Start by defining the endpoint and response direction. An agonist response, inhibition readout, binding signal, and relative-potency comparison can use similar sigmoid models while answering different questions.

Preserve individual wells and replicate identities before deriving normalized responses. A plotted mean without its source wells can hide plate position, exclusions, and variability.

  1. Retain raw response, plate position, sample, concentration, replicate, and control labels.
  2. Apply blank correction or normalization according to the method, recording the exact transformation.
  3. Inspect concentration coverage and confirm that the fitted midpoint is supported by observations on both sides.
  4. Fit the candidate model with a documented equation and constraints.
  5. Review residuals, parameter uncertainty, curve shape, and assay-specific suitability checks.
  6. Report the estimate with units, model version, exclusions, diagnostics, and fingerprints for the source data file and analysis configuration.

Choose the model with scientific and diagnostic evidence

A 4PL is a common starting point for symmetric sigmoidal behavior. A 5PL adds asymmetry, but the extra parameter also needs enough informative data. Constraining an asymptote may stabilize a fit when scientifically justified; it also changes the estimand and must be visible.

R² alone is not a model-selection rule. Residual patterns, parameter precision, convergence, concentration coverage, weighting assumptions, and performance on known samples provide more useful evidence.

QuestionEvidence to inspect
Is the midpoint interpolated?Observed concentrations and fitted midpoint
Is asymmetry supported?Residual pattern, parameter uncertainty, repeated-run behavior
Are constraints justified?Method rationale and performance against unconstrained fits
Is weighting needed?Variance across the response or concentration range

Make the result reproducible

Record whether concentrations were fitted on the raw or logarithmic scale, how zero-dose controls were handled, whether responses were normalized, and which definition of EC50 or IC50 was reported.

The report should connect the plotted curve to model parameters, well-level observations, exclusions with reasons, suitability results, the software version, and the analysis settings used to generate it.

Frequently asked questions

Should zero concentration be included on a logarithmic x-axis?

Logarithms of zero are undefined. Preserve zero-dose controls as observations, but display or model them using an explicitly documented method rather than silently replacing zero with an arbitrary small concentration.

Is the model with the highest R² always best?

No. R² may improve when parameters are added even when the added flexibility is unstable or scientifically unhelpful. Inspect residuals, uncertainty, coverage, convergence, and repeated-run performance.

Primary references