476
N. Ya. Golovenko et al.
The approach outlined above has already been shown to be highly efficient for
solving various “structure-activity” tasks [23, 27, 30, 57–59]. SiRMS methodology
does not have the restrictions of such well-known and widely used approaches as
CoMFA, CoMSIA, and HASL, in which the application is limited to a structurally
homogeneous set of molecules only. SiRMS approaches also do not have the disadvantages of the HQSAR approach [36, 52] that are related to the ambiguity of the
descriptors system formation.
14.4.1.1 QSPR Prediction of the Drugs Bioavailability on the Base
of the SiRMS Approach
Influence of the molecular structure of drug API on absorption was investigated using SiRMS QSAR approach [26, 60]. Data for human oral bioavailability were obtained from the literature and an internal database. SMILES strings were retrieved
from the World Drug Index (WDI, Derwent Publishers, London) or created manually. Finally, 628 structures with SMILES, generic name, and bioavailability value
were obtained. Any compounds whose bioavailability is strongly affected by the
dose and formulation was excluded from the dataset. Random forest method [61,
62] was used for the development of QSPR classification models. All compounds
were divided into three (high, medium and low bioavailability, Table 14.1) or two
(high and acceptable, Table 14.2) classes. To determine whether the classes limits
shift affects the quality of predictive models and how it affects, were considered
different limits of the classes, as shown in Table 14.1.
As seen from the results presented above, misclassification error is high and
hence predictive ability of models is rather low. Model II has the best prediction
performance, especially for classes 1 and 2. Regression model with low predictive
ability (R
2
ооb
= 0.29) was also obtained using this dataset of 628 compounds.
Binary classification model has better predictive ability (Table 14.2).
Table 14.2 shows that the shift of models boundaries reduces accuracy of their
prediction. Consequently, in order to identify the causes of this phenomenon, we
conducted a thorough analysis of medicinal agents by all relevant models (IV—VІ).
The analysis of the results showed the good concordance between models IV-VI. It
have been proved that prediction error for the class 1 (bioavailability of 90–100, or
80–100, or 70–100) is associated with the active participation of specific transporters during absorption. For class 2, insignificant participation of transporters in the
absorption of medicinal agents has been observed, but there is an intensive metabolism of compounds and their binding to targets (furthermore, the most of medicinal
agents have multiple targets, binding with which is strong and durable).
Taking into account intersection of models and variation of bioavailability in
some range of values, we have introduced to the obtained model a confidence interval within 5 % on either side of the chosen threshold value (Fig. 14.3).
Thus, the boundary between the classes has a certain thickness. Errors of classification shall be specified in the case when the molecules are predicted to be in the
opposite class beyond the boundary limits.
N. Ya. Golovenko et al.
The approach outlined above has already been shown to be highly efficient for
solving various “structure-activity” tasks [23, 27, 30, 57–59]. SiRMS methodology
does not have the restrictions of such well-known and widely used approaches as
CoMFA, CoMSIA, and HASL, in which the application is limited to a structurally
homogeneous set of molecules only. SiRMS approaches also do not have the disadvantages of the HQSAR approach [36, 52] that are related to the ambiguity of the
descriptors system formation.
14.4.1.1 QSPR Prediction of the Drugs Bioavailability on the Base
of the SiRMS Approach
Influence of the molecular structure of drug API on absorption was investigated using SiRMS QSAR approach [26, 60]. Data for human oral bioavailability were obtained from the literature and an internal database. SMILES strings were retrieved
from the World Drug Index (WDI, Derwent Publishers, London) or created manually. Finally, 628 structures with SMILES, generic name, and bioavailability value
were obtained. Any compounds whose bioavailability is strongly affected by the
dose and formulation was excluded from the dataset. Random forest method [61,
62] was used for the development of QSPR classification models. All compounds
were divided into three (high, medium and low bioavailability, Table 14.1) or two
(high and acceptable, Table 14.2) classes. To determine whether the classes limits
shift affects the quality of predictive models and how it affects, were considered
different limits of the classes, as shown in Table 14.1.
As seen from the results presented above, misclassification error is high and
hence predictive ability of models is rather low. Model II has the best prediction
performance, especially for classes 1 and 2. Regression model with low predictive
ability (R
2
ооb
= 0.29) was also obtained using this dataset of 628 compounds.
Binary classification model has better predictive ability (Table 14.2).
Table 14.2 shows that the shift of models boundaries reduces accuracy of their
prediction. Consequently, in order to identify the causes of this phenomenon, we
conducted a thorough analysis of medicinal agents by all relevant models (IV—VІ).
The analysis of the results showed the good concordance between models IV-VI. It
have been proved that prediction error for the class 1 (bioavailability of 90–100, or
80–100, or 70–100) is associated with the active participation of specific transporters during absorption. For class 2, insignificant participation of transporters in the
absorption of medicinal agents has been observed, but there is an intensive metabolism of compounds and their binding to targets (furthermore, the most of medicinal
agents have multiple targets, binding with which is strong and durable).
Taking into account intersection of models and variation of bioavailability in
some range of values, we have introduced to the obtained model a confidence interval within 5 % on either side of the chosen threshold value (Fig. 14.3).
Thus, the boundary between the classes has a certain thickness. Errors of classification shall be specified in the case when the molecules are predicted to be in the
opposite class beyond the boundary limits.