362
A. Zakharov and A. Lagunin
as QSAR equation is much more complicated. However, there are several computer
programs designed to solve those problems [82].
Another limitation is the difficulty of combining biological data generated in real
time with computer predictions. Nevertheless, there are few successful examples
of this approach, such as association of structure metabolites predictions with the
spectra of liquid chromatography in mass spectrometry [83, 84].
It is considered to be the fundamental concept that the computer prediction of
toxicity applied to the analysis of preclinical drug compounds, in fact, is a prediction of the prediction. It is necessary to take into account that most preclinical
parameters (e.g., carcinogenicity, genetic toxicity and teratogenicity) are predictions of human toxicity, which help to establish the safety of drugs before their
clinical trials. Thus, the creation of models based on other models, can only result
in uncertainty. In reality, (Q)SAR is a theoretical analysis, based on the modeling
of the chemical space and data from human toxicity models [1]. Unfortunately,
uncertainty is still inevitable, because another layer of the modeling is added for the
safety assessment. QSAR predictions have to be based on the same type of features
among many important pieces of information (e.g., duration and level of exposure,
confounding factors, and a risk/benefit ratio) in the general risk analysis. An applicability domain of the model is also the main criterion and restricting factor to use
(Q)SAR models in toxicology and pharmacology [85, 86]. If a compound does not
fall into an applicability domain of the model during in silico screening the prediction is considered to be incorrect. At the same time, it can be expected that the development of new chemical entities requires moving towards new chemical spaces,
since they are created for new therapeutic targets. Thus, the domain of applicability
for new QSAR models is needed to be extended for new molecules. This will lead to
overcoming limitations and thus provide more accurate and acceptable predictions.
Predictions of a specific toxicity (e.g., carcinogenicity) based on QSAR models
associated with the alerts classification schemes (e.g. Ashby-Tennant alerts) or expert rules, also have their own limitations. A significant part of used drugs shows a
positive result in the rodent’s carcinogenicity and negative results for genotoxicity
[87]. It leads to the dilemma how to predict non genotoxic carcinogens, can it be
based either on the structure formula or on the structural alerts associated with the
inducing of DNA damage? Therefore, the developments, which are devoted to this
problem, and the predictions of epigenetic mechanisms with a carcinogens action
have the primary importance. In addition, recently it has been proposed that there is
a new research area for computer predictions of carcinogenicity: creating computer
models to predict carcinogens acting through inhibition of protein kinase networks.
The expert system based on rules has limitations in an ability to identify structural alerts, which lead to the manifestation of activity and does not have built-in
rules for the estimation of compounds, which includes two or more structural alert
or deactivating fragments in the molecule. Moreover, the “negative” prediction of
these programs means that nothing has been found and prediction could not indicate
the loss of toxicity. Finally, there is a human factor underlying this approach. It is
represented by the consensus opinion or expert opinion which are prone to subjectivity and can result in incorrect or inaccurate prediction rules [88].
A. Zakharov and A. Lagunin
as QSAR equation is much more complicated. However, there are several computer
programs designed to solve those problems [82].
Another limitation is the difficulty of combining biological data generated in real
time with computer predictions. Nevertheless, there are few successful examples
of this approach, such as association of structure metabolites predictions with the
spectra of liquid chromatography in mass spectrometry [83, 84].
It is considered to be the fundamental concept that the computer prediction of
toxicity applied to the analysis of preclinical drug compounds, in fact, is a prediction of the prediction. It is necessary to take into account that most preclinical
parameters (e.g., carcinogenicity, genetic toxicity and teratogenicity) are predictions of human toxicity, which help to establish the safety of drugs before their
clinical trials. Thus, the creation of models based on other models, can only result
in uncertainty. In reality, (Q)SAR is a theoretical analysis, based on the modeling
of the chemical space and data from human toxicity models [1]. Unfortunately,
uncertainty is still inevitable, because another layer of the modeling is added for the
safety assessment. QSAR predictions have to be based on the same type of features
among many important pieces of information (e.g., duration and level of exposure,
confounding factors, and a risk/benefit ratio) in the general risk analysis. An applicability domain of the model is also the main criterion and restricting factor to use
(Q)SAR models in toxicology and pharmacology [85, 86]. If a compound does not
fall into an applicability domain of the model during in silico screening the prediction is considered to be incorrect. At the same time, it can be expected that the development of new chemical entities requires moving towards new chemical spaces,
since they are created for new therapeutic targets. Thus, the domain of applicability
for new QSAR models is needed to be extended for new molecules. This will lead to
overcoming limitations and thus provide more accurate and acceptable predictions.
Predictions of a specific toxicity (e.g., carcinogenicity) based on QSAR models
associated with the alerts classification schemes (e.g. Ashby-Tennant alerts) or expert rules, also have their own limitations. A significant part of used drugs shows a
positive result in the rodent’s carcinogenicity and negative results for genotoxicity
[87]. It leads to the dilemma how to predict non genotoxic carcinogens, can it be
based either on the structure formula or on the structural alerts associated with the
inducing of DNA damage? Therefore, the developments, which are devoted to this
problem, and the predictions of epigenetic mechanisms with a carcinogens action
have the primary importance. In addition, recently it has been proposed that there is
a new research area for computer predictions of carcinogenicity: creating computer
models to predict carcinogens acting through inhibition of protein kinase networks.
The expert system based on rules has limitations in an ability to identify structural alerts, which lead to the manifestation of activity and does not have built-in
rules for the estimation of compounds, which includes two or more structural alert
or deactivating fragments in the molecule. Moreover, the “negative” prediction of
these programs means that nothing has been found and prediction could not indicate
the loss of toxicity. Finally, there is a human factor underlying this approach. It is
represented by the consensus opinion or expert opinion which are prone to subjectivity and can result in incorrect or inaccurate prediction rules [88].