TrueGate Bio · Analytics Case Work

Relative potency: is the curve parallel?

A four-parameter logistic analysis of a cell-based potency assay, and why a non-parallel sample has no reportable relative potency.

Synthetic data, built to demonstrate the method

The question

A cell-based potency assay reports relative potency by comparing each sample's dose-response curve with a reference standard. Two samples both return a relative potency number from the software. Can both be reported?

The analysis

2026-09-25T02:57:46.889215 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

Relative potency only means something when the test curve is the reference curve shifted along the dose axis: same shape, same slope, same response range. That is the parallelism assumption. Each curve was fitted with a four-parameter logistic model, and parallelism was judged by equivalence: the 90% confidence intervals for the slope ratio and the response-range ratio must sit inside 0.80 to 1.25, limits set before testing.

SampleSlope ratio (90% CI)Response-range ratio (90% CI)ParallelismRelative potency (90% CI)
Sample A1.01 (0.90 to 1.14)1.00 (0.96 to 1.05)Parallel89% (80 to 100%)
Sample B0.79 (0.70 to 0.91)0.64 (0.60 to 0.68)Not parallelNot reportable
Why it matters. Sample A is parallel and its relative potency is meaningful. Sample B has a shallower slope and a lower maximum response, which is what a partly degraded or differently acting product can look like. Software will still compute a number for it, but that number compares two different curves and should not be reported.

What to do

  • Make parallelism a system suitability requirement with pre-set equivalence limits, not an optional check.
  • Treat a non-parallel result as a signal about the product, and investigate it rather than repeating the assay until it passes.
  • Report relative potency with its confidence interval, not as a single number.

How it was done

Synthetic dose-response data, eight doses in triplicate. Four-parameter logistic fits by nonlinear least squares; confidence intervals from 400 residual bootstrap resamples; equivalence limits of 0.80 to 1.25, following the equivalence approach described in USP chapter 1032. Download the data (CSV).

Back to all work   Discuss a similar problem