468
N. Ya. Golovenko et al.
program). Moreover, “the rule of five” was developed when information about drug
transporters was very limited and so it described links between physico–chemical
properties of drugs and their bioavailability based on API absorption determined by
simple diffusion.
In addition to the molecular properties discussed by Lipinski, other properties
have been discussed in regard to oral bioavailability. Since then, numerous classification and regression models for the predictions of absorption were reported by
applying a variety of statistical and machine-learning approaches, which include
multiple linear regression (MLR), nonlinear regression, partial least squares (PLS)
regression, linear discriminant analysis, classification and regression trees, artificial
neural networks (ANN), support vector machines (SVM), and so forth. Because
of the fact that many factors are related to intestinal absorption, many physicochemical descriptors were introduced into the prediction of absorption and bioavailability, such as polar surface area (PSA), partition coefficients, molecular size,
hydrogen-bonding descriptors, topological descriptors, and even quantum chemical
descriptors. Predictivity of models bioavailability is by prediction of estimated independent compounds. To be useful in drug development, models should ultimately
be developed to predict unknown compounds.
In subsequent work, other researchers have introduced rules-of-thumb which can
increase the chances of drug compounds being well absorbed. In 2002, Veber and
colleagues reported studies on rat bioavailability data for 1100 drug candidates [27].
They proposed two other descriptors and suggested that compounds which meet
only two criteria of (a) 10 or fewer rotatable bonds and (b) polar surface area equal
to or less 140 Å
2
(or 12 or fewer H-bond donors and acceptors) will have a high
probability of good oral bioavailability in rat. In 2004, Lu and colleagues investigated the relationship between the number of rotatable bonds and PSA for rat oral bioavailability using 434 molecules [28]. Compared to Veber’s work, authors reported
that the prediction results were dependent on the calculation methods. In 2007, Hou
and colleagues [29] collected a dataset of 773 compounds with experimental human
oral bioavailability values. They showed that the percentages of compounds meeting the criteria based on molecular properties does not distinguish compounds with
poor oral bioavailability from those with acceptable values. A dataset of intestinal
absorption was also examined and compared with that of oral bioavailability. The
performance of these rules based on molecular properties in the prediction of intestinal absorption is obviously much better than that of oral bioavailability in term of
false positive rate, and, therefore, the applications of the “rule-based” approaches
on the prediction of human bioavailability should be very cautious. Then, Veber’s
rules were used for the entire dataset to see if these rules could be applied for the
prediction of human oral bioavailability. Afterwards, the correlations between several important molecular descriptors and human oral bioavailability were examined.
They conjectured that there are no simple rules based on molecular descriptors that
can be used to predict human oral bioavailability truly well compared to the rules
based on analyzing rat oral bioavailability data [29]. However, the value of developed models and all the conclusions mentioned above is significantly decreased
N. Ya. Golovenko et al.
program). Moreover, “the rule of five” was developed when information about drug
transporters was very limited and so it described links between physico–chemical
properties of drugs and their bioavailability based on API absorption determined by
simple diffusion.
In addition to the molecular properties discussed by Lipinski, other properties
have been discussed in regard to oral bioavailability. Since then, numerous classification and regression models for the predictions of absorption were reported by
applying a variety of statistical and machine-learning approaches, which include
multiple linear regression (MLR), nonlinear regression, partial least squares (PLS)
regression, linear discriminant analysis, classification and regression trees, artificial
neural networks (ANN), support vector machines (SVM), and so forth. Because
of the fact that many factors are related to intestinal absorption, many physicochemical descriptors were introduced into the prediction of absorption and bioavailability, such as polar surface area (PSA), partition coefficients, molecular size,
hydrogen-bonding descriptors, topological descriptors, and even quantum chemical
descriptors. Predictivity of models bioavailability is by prediction of estimated independent compounds. To be useful in drug development, models should ultimately
be developed to predict unknown compounds.
In subsequent work, other researchers have introduced rules-of-thumb which can
increase the chances of drug compounds being well absorbed. In 2002, Veber and
colleagues reported studies on rat bioavailability data for 1100 drug candidates [27].
They proposed two other descriptors and suggested that compounds which meet
only two criteria of (a) 10 or fewer rotatable bonds and (b) polar surface area equal
to or less 140 Å
2
(or 12 or fewer H-bond donors and acceptors) will have a high
probability of good oral bioavailability in rat. In 2004, Lu and colleagues investigated the relationship between the number of rotatable bonds and PSA for rat oral bioavailability using 434 molecules [28]. Compared to Veber’s work, authors reported
that the prediction results were dependent on the calculation methods. In 2007, Hou
and colleagues [29] collected a dataset of 773 compounds with experimental human
oral bioavailability values. They showed that the percentages of compounds meeting the criteria based on molecular properties does not distinguish compounds with
poor oral bioavailability from those with acceptable values. A dataset of intestinal
absorption was also examined and compared with that of oral bioavailability. The
performance of these rules based on molecular properties in the prediction of intestinal absorption is obviously much better than that of oral bioavailability in term of
false positive rate, and, therefore, the applications of the “rule-based” approaches
on the prediction of human bioavailability should be very cautious. Then, Veber’s
rules were used for the entire dataset to see if these rules could be applied for the
prediction of human oral bioavailability. Afterwards, the correlations between several important molecular descriptors and human oral bioavailability were examined.
They conjectured that there are no simple rules based on molecular descriptors that
can be used to predict human oral bioavailability truly well compared to the rules
based on analyzing rat oral bioavailability data [29]. However, the value of developed models and all the conclusions mentioned above is significantly decreased