11 Applicability Domain: Towards a More Formal …
219
Fig. 11.2 A confident decision can be made based on a model prediction, when the prediction
is valid (in the Applicability Domain), reliable (in the Reliability Domain) and decisive (in the
Decidability Domain). If all three criteria are met, the prediction is said to be in the Decision
Domain
11.2.1 Applicability
Prior to achieving any prediction, it is critical to check if it is valid to apply the
model for the compound being predicted; in other words, we check that the model’s
specifications are compliant with the intended use case. These steps can be seen as
a checklist with questions like:
• Is the chemical class of the query compound supported by the model? An example
of such a specification could be “accept organic compounds but exclude polymers,
proteins and inorganic molecules”. There are many possible causes for the limitations of a model. For instance, for some classes of compounds, the model could be
unable to compute some of the required descriptors or the model’s internal chemical representation could be limited to a specific class of compound, e.g. organic
molecules only due to the complex bonds in inorganic molecules, the size of the
proteins structures or the repeated motifs in polymers (Fig. 11.3). Such specifications are statically attached to the model irrespective of the predicted compound
or the training dataset.
• Are the values of the descriptors of the predicted compound within the range of
values the model has been trained or designed for? Here the descriptor space is
used to restrain the applicability domain of the model to ensure that the query
structure does not exhibit unusual properties for which the model has not been
trained. Different existing AD methods are based on this criteria amongst which
the most popular are:
– Range Box: This naïve approach defines the boundaries of the applicability
domain as the range between the minimum and the maximum values, for each
individual descriptor variable as observed in the training data. If any descriptor of
the predicted compound has a value outside the interval [minimum-maximum]
219
Fig. 11.2 A confident decision can be made based on a model prediction, when the prediction
is valid (in the Applicability Domain), reliable (in the Reliability Domain) and decisive (in the
Decidability Domain). If all three criteria are met, the prediction is said to be in the Decision
Domain
11.2.1 Applicability
Prior to achieving any prediction, it is critical to check if it is valid to apply the
model for the compound being predicted; in other words, we check that the model’s
specifications are compliant with the intended use case. These steps can be seen as
a checklist with questions like:
• Is the chemical class of the query compound supported by the model? An example
of such a specification could be “accept organic compounds but exclude polymers,
proteins and inorganic molecules”. There are many possible causes for the limitations of a model. For instance, for some classes of compounds, the model could be
unable to compute some of the required descriptors or the model’s internal chemical representation could be limited to a specific class of compound, e.g. organic
molecules only due to the complex bonds in inorganic molecules, the size of the
proteins structures or the repeated motifs in polymers (Fig. 11.3). Such specifications are statically attached to the model irrespective of the predicted compound
or the training dataset.
• Are the values of the descriptors of the predicted compound within the range of
values the model has been trained or designed for? Here the descriptor space is
used to restrain the applicability domain of the model to ensure that the query
structure does not exhibit unusual properties for which the model has not been
trained. Different existing AD methods are based on this criteria amongst which
the most popular are:
– Range Box: This naïve approach defines the boundaries of the applicability
domain as the range between the minimum and the maximum values, for each
individual descriptor variable as observed in the training data. If any descriptor of
the predicted compound has a value outside the interval [minimum-maximum]