Information Technology Reference
In-Depth Information
classifi cation of discrete values, calculation of the correlation coeffi cient
for tree predictions of unseen test samples is not possible. The average
confi dence level of predictions and the percentage of correct predictions
can be employed to assess the predictability of decision trees. Decision
trees are prone to errors in classifi cation when many classes and a relatively
small number of samples in a data set are used for tree construction.
5.3.3 Examples
Decision trees have been used to classify CYP450 3A4 inhibitors and
non-inhibitors (Choi, 2009); dopamine, serotonine, and dual dopamine-
serotonine antagonists (Kim et al., 2006); hERG channel blockers (Gepp
and Hutter, 2006); acetylcholinesterase inhibitors (Lv and Xue, 2010); as
an aid in intravenous formulation development (Lee et al., 2003); and to
generate an expert system for the identifi cation of fi lm coating (Rowe and
Upjohn, 1993).
Several illustrative examples of decision tree implementation are now
described in more detail.
Example 1
(Branchu et al., 2007)
Physicochemical data (pK
a
values; molecular weight, volume, and
surface area; numbers of hydrogen bond donors and acceptors;
total and percent polar surface area; Lipinski score; solubility in μg/
ml in human intestinal fl uid or other aqueous media; dose in mg,
logarithm of dose number; logarithm of the experimental or calculated
distribution coeffi cient between octanol and buffer in the small
intestinal pH range; melting point) for a set of potentially poorly
soluble compounds were analyzed in relation to suitable formulations
for these compounds. Physical chemistry was found to be a key
determinant of formulation class expressed in terms of conventional,
solid dispersion, lipidic/surfactant, and crystalline nanoparticle
systems. This relationship was used to build a decision-support tool
aimed to guide formulation selection for poorly soluble compounds
during product development. Prior to decision tree building, data
were normalized to a zero to unity range. The decision tree underwent
cross-validation, in order to check its performance.
Search WWH ::
Custom Search