How expert rule-based and statistical (Q)SAR approaches complement each other in mutagenic impurity assessment.
Rule-based and statistical models address the same endpoint from different scientific perspectives. Their value lies in complementarity rather than in treating either result as a standalone truth.
Rule-based strengths
Transparent alerts can connect structural features with known electrophilic, DNA-reactive or metabolically activated mechanisms.
Statistical strengths
Data-driven models can capture patterns that are not easily represented by a fixed alert library and can exploit similarity to experimental chemical space.
Discordant predictions
Differences between models should trigger scientific interpretation, including alert relevance, analogues and domain support.
Regulatory conclusion
The final assessment is a weight-of-evidence conclusion, not simply a vote between two software outputs.
Frequently asked questions
Which approach is better?
They are complementary; neither is universally superior.
What if the two models disagree?
Investigate the reason for discordance and use expert review.
Does statistical prediction need an applicability domain?
Yes.
Are structural alerts mechanistically useful?
They can be, especially when the alert is chemically plausible and relevant to the query structure.
