218
T. Hanser et al.
• Are there given structures for which the model should not be used?
• How do we define the boundaries of the chemical space?
• How do we define the reliability of a prediction?
• When is a given prediction reliable enough?
• Is the prediction conclusive, i.e. is the outcome likely?
One of the main sources of confusion is probably a natural desire to compile
all these different questions into one: “Is the query compound inside or outside the
applicability domain?”, in other words, into a closed yes/no question. This compilation has several drawbacks; first, it is very difficult to combine information of a
different nature and express the resulting concept in a single metric, secondly, the
resulting AD implementation is less interpretable since it becomes very challenging,
a posteriori, to disentangle the merged information. Finally, the assessment of the
confidence in a prediction follows a chronology that is not captured in a single closed
question. The importance of such a timeline will be discussed later. Another risk is
to address only a subset of these questions which is unfortunately the case for many
AD methodologies.
11.2 Decision Domain
A more holistic vision of AD becomes apparent when looking from the broader
decision perspective; this approach was introduced by Hanser et al. [16]. Indeed,
QSAR models are most useful if they provide sufficient support to the user to enable
a confident decision to be made. If this is not the case, then the model offers little
value.
From such a perspective, we can identify three well-defined concerns, each
addressing a specific AD aspect along with a clear chronology that leads to a threestaged approach:
1. Applicability: Can the model be applied to make a prediction for my query?
2. Reliability: Is the resulting prediction reliable enough for the intended use case?
3. Decidability: Is the outcome of the prediction decisive (unequivocal)?
Together, these requirements define the scope for a confident decision based on a
prediction (Fig. 11.2) called the decision domain (DD).
The decision domain can be defined as follows:
The Decision Domain is the scope within which it is possible to make a confident decision
based on a valid, reliable and decisive prediction.
By separating the three key concepts embedded in a prediction and addressing
them individually, we can use the appropriate methodologies for assessing these
aspects. This new perspective results in a more formal framework that supports
the decision-making process. Existing AD methodologies can be mapped onto this
framework, and we will use this structure to introduce these approaches at their
corresponding level.
T. Hanser et al.
• Are there given structures for which the model should not be used?
• How do we define the boundaries of the chemical space?
• How do we define the reliability of a prediction?
• When is a given prediction reliable enough?
• Is the prediction conclusive, i.e. is the outcome likely?
One of the main sources of confusion is probably a natural desire to compile
all these different questions into one: “Is the query compound inside or outside the
applicability domain?”, in other words, into a closed yes/no question. This compilation has several drawbacks; first, it is very difficult to combine information of a
different nature and express the resulting concept in a single metric, secondly, the
resulting AD implementation is less interpretable since it becomes very challenging,
a posteriori, to disentangle the merged information. Finally, the assessment of the
confidence in a prediction follows a chronology that is not captured in a single closed
question. The importance of such a timeline will be discussed later. Another risk is
to address only a subset of these questions which is unfortunately the case for many
AD methodologies.
11.2 Decision Domain
A more holistic vision of AD becomes apparent when looking from the broader
decision perspective; this approach was introduced by Hanser et al. [16]. Indeed,
QSAR models are most useful if they provide sufficient support to the user to enable
a confident decision to be made. If this is not the case, then the model offers little
value.
From such a perspective, we can identify three well-defined concerns, each
addressing a specific AD aspect along with a clear chronology that leads to a threestaged approach:
1. Applicability: Can the model be applied to make a prediction for my query?
2. Reliability: Is the resulting prediction reliable enough for the intended use case?
3. Decidability: Is the outcome of the prediction decisive (unequivocal)?
Together, these requirements define the scope for a confident decision based on a
prediction (Fig. 11.2) called the decision domain (DD).
The decision domain can be defined as follows:
The Decision Domain is the scope within which it is possible to make a confident decision
based on a valid, reliable and decisive prediction.
By separating the three key concepts embedded in a prediction and addressing
them individually, we can use the appropriate methodologies for assessing these
aspects. This new perspective results in a more formal framework that supports
the decision-making process. Existing AD methodologies can be mapped onto this
framework, and we will use this structure to introduce these approaches at their
corresponding level.