Applicability domain gives a prediction context
A (Quantitative) Structure-Activity Relationship (Q)SAR result needs context before it can support a regulatory conclusion.
For mutagenicity assessment under ICH M7, one part of that context is the applicability domain (AD): whether the query compound is sufficiently represented by the chemistry used to develop the model. This matters because a statistical model performs most reliably when it is making predictions within chemical space supported by its training data.
The AD indicates whether the query compound is sufficiently represented within the model’s training set to support a prediction. This information is meaningful because it helps define the context in which the prediction should be interpreted and highlights when additional review or supporting evidence may be warranted. A defined applicability domain is also one of the OECD principles for the validation of (Q)SAR models for regulatory purposes.
Chemical coverage matters
Statistical models learn associations between chemical structure and experimental outcomes from curated training data. The composition of that data therefore places practical boundaries around what the model can predict confidently. There are multiple methods for assessing a model’s applicability domain. For example, a prediction may be considered outside the applicability domain when there are no relevant structural features represented by the model or when there are insufficiently analogous structures within the training set to support the query.
Without an appropriate AD definition, a prediction may be communicated with more certainty than the underlying evidence supports. In the case of negative predictions generated from models lacking sufficiently robust AD definitions, there is a risk of interpreting the result as stronger evidence for an absence of mutagenic potential than the underlying model support warrants. Therefore, a clearly characterized applicability domain helps to make uncertainty visible and appropriately communicate the result
Increasing chemical domain coverage is an important goal of model development. By expanding the number of training compounds, underrepresented chemistry may be captured. However, identifying gaps in chemical space can help guide targeted expansion of the applicability domain for specific chemotypes. Regardless of approach, new data must be experimentally reliable, chemically correct, and relevant to the endpoint. As such, domain expansion is valuable when it increases relevant chemical representation while preserving the scientific integrity and predictive performance of the model.
Applicability domain and ICH M7 assessment
ICH M7 uses two complementary (Q)SAR methodologies for bacterial mutagenicity assessment: an expert rule-based methodology and a statistical-based methodology.
The statistical prediction is therefore considered alongside the complementary methodology. Applicability-domain information provides context for understanding the model support for prediction. Where a statistical prediction is out-of-domain, additional relevant evidence can include any matched alerts from a complementary expert rule-based system, an assessment of potentially reactive features, mechanistic information, suitable analogs and if available any supporting experimental data. The expert review then determines whether those lines of evidence support a scientifically justified conclusion.
This is where transparency in QSAR becomes important. A prediction is easier to defend when the assessor can examine the evidence behind it rather than treating the model output as a black box.
Expanding domain without weakening the model
As new and increasingly diverse chemistries are developed and additional high-quality bacterial mutagenicity data become available, statistical models can incorporate chemistry that was previously underrepresented. Done carefully, this can reduce the frequency of out-of-domain predictions, while also preserving or improving predictivity.
For Leadscope, balancing chemical domain coverage with model performance is central to the development of statistical models, including the bacterial mutation models. Identifying gaps in model coverage can inform targeted expansion of relevant chemical space, while careful data selection and model validation help ensure that broader applicability does not come at the expense of predictive performance.
Applicability informs interpretation
Applicability domain is critical because it prevents QSAR predictions from being interpreted without its scientific boundaries.
For an ICH M7 assessment, knowing that a compound is outside the statistical model’s domain changes the evidentiary question. It directs attention to chemistry, complementary prediction methodology, mechanistic support, and other available evidence.
In regulatory toxicology, applicability domain provides important context for understanding the computational support for a prediction and interpreting that prediction alongside other available evidence.
Mehr erfahren
Leadscope supports ICH M7 assessments through industry leading complementary statistical and expert rule-based methodologies, curated toxicology data and tools for examining the evidence associated with prediction.
Explore how Leadscope supports transparent evaluation of QSAR predictions and their applicability across regulatory safety assessments.
