220
T. Hanser et al.
Fig. 11.3 A model may only accept some classes of query compounds. For instance, polymers,
proteins and organometallic structures may be too complex or too large for the models’ internal
representation and should be excluded from the Applicability Domain
Fig. 11.4 Min-max
descriptor ranges observed in
the training data (blue) define
the acceptable values of the
query structure descriptors in
the form of a hyper-box
(here a rectangle in the case
of two dimensions)
of the corresponding variable in the training data, then the model is not applicable, and this compound is labelled as “Outside the Applicability Domain”
(Fig. 11.4). This approach bears an important drawback as it allows regions
containing very few training examples inside the hyper-box defined by these
ranges. As the number of descriptors (dimension of the descriptor space) use
by the model increases, these regions become rapidly predominant due to the
dimensionality curse [17].
– Convex Hull: This approach is very similar to the range box method with the
difference that instead of a hyper-box defined in the descriptor space, the convex
hull is the minimum surface that wraps the training data, defining a more precise
T. Hanser et al.
Fig. 11.3 A model may only accept some classes of query compounds. For instance, polymers,
proteins and organometallic structures may be too complex or too large for the models’ internal
representation and should be excluded from the Applicability Domain
Fig. 11.4 Min-max
descriptor ranges observed in
the training data (blue) define
the acceptable values of the
query structure descriptors in
the form of a hyper-box
(here a rectangle in the case
of two dimensions)
of the corresponding variable in the training data, then the model is not applicable, and this compound is labelled as “Outside the Applicability Domain”
(Fig. 11.4). This approach bears an important drawback as it allows regions
containing very few training examples inside the hyper-box defined by these
ranges. As the number of descriptors (dimension of the descriptor space) use
by the model increases, these regions become rapidly predominant due to the
dimensionality curse [17].
– Convex Hull: This approach is very similar to the range box method with the
difference that instead of a hyper-box defined in the descriptor space, the convex
hull is the minimum surface that wraps the training data, defining a more precise