The statistical likelihood for failures now needs to be estimated and a component variance analysis performed to estimate
the contribution of each component. Then, focus can be put on how to set limits on post-validation activities from understanding
the potential impact on the likelihood for all cases (1A-2B) to occur. The situation can be most effectively improved by having
primarily in mind patient safety, dosing, and regulatory expectations, but secondarily also the firm as the need to pass specifications
and stay profitable is important. Similar to propagation-of-error calculations, an estimate for the sampling variance (batch-uniformity,
stability, protein adsorption losses, etc.), could allow immediate estimation of the actual (true) process performance for
potency by simply solving for it from Equation 1 (V = variance).
The (hypothetical) historical results are presented in Table 2. The estimated probabilities for cases 1A and 2A, and the measurement
errors ("invisible" component of SPC) are given in Table 3 along with probability estimates for the worst-case scenario when
AMT results would reveal that the receiving laboratory will test at a bias at the acceptable AMT protocol limit of ±1.5%.
Although simplified, several observations can be made from the data in Figure 1 and Table 2. Most important, the overall process
performance is out of a desirable ±3 s.d. SPC state. As said earlier, this is the "visible" SPC trigger that warrants action
when limits are exceeded. The test system has recently (last n = 60) yielded higher (2.0%) results for the assay control.
The assay control is the same molecule in a highly similar matrix as the test sample and both are run simultaneously in each
The "visible" process mean can therefore be expected to be about 2% higher. Looking chronologically at the assay control may
show the root cause, that is, which assay component was changed and caused this difference over time. Alternatively, several
smaller changes could have added up to this difference of 2.0%. For example, the reference standard may unknowingly be unstable
even when frozen and may have lost 2.0% potency over time, providing proportionally higher (2.0%) results to the assay control
and test samples.
Although the 2.0% expected difference in process data may in reality be buried within several small sampling or production
changes, it nevertheless constitutes a 2.0% bias for testing from the time that specifications were set to match clinical
data, and the then-existing process variance (PV) and AMV.
The estimate for the overall sampling process during production (from PV) is used to estimate the actual variance in process
performance. However, if small- or large-scale process studies to estimate sampling variance were not done well or are no
longer representative, this estimate may not be sufficiently accurate to provide a good estimate for the actual PV (using
Equation 1). In the hypothetical example, the estimated true PV (2.0%) is smaller than the estimated sampling variance (2.3%),
and smaller than the assay variance (3.0%) and always smaller than the overall "visible" process variance (4.3%). This may
have been why the specifications had been set to 90–110 units in the first place. Often, the assay and sampling variance could
indeed be greater than the actual process variance for downstream in-process potency testing for biopharmaceuticals because
of various reasons. Some examples are listed below:
- Protein adsorption losses (sampling, testing).
- Sampling procedures lacking detail where needed.
- Inappropriate sample handling before testing.
- Poorly set test system suitability criteria.
- Insufficiently monitored and controlled assay reproducibility.
- Poorly developed or optimized test method.
- Poorly written AMV protocol.