361
11 Computational Toxicology in Drug Discovery: Opportunities and Limitations
11.5 A Critical Assessment of Computational Approaches
for Toxicity Prediction
For a better understanding of the possible toxicity prediction it is necessary to pay
attention to a number of limitations that may occur using modern computer tools
for creation of (Q)SAR models. The major limitation factors are the guarantee of
high-quality experimental data used for creation of the training sets and understanding of what is exactly modeled by the user. When mistakes occur (e.g., an incorrect
structure of the molecule or incorrect data from toxicological studies) in the training
set, it leads to the wrong model, which provides incorrect predictions. Therefore,
considerable efforts should be made for the appropriate high-quality selection of experimental data, which will be used for creation of the model. Some recommendations for data curation in cheminformatics and QSAR modeling were published by
Fourches and co-authors [81]. Since the concept of the “most suitable” and “quality” is subjective, even among experts, determination of the quality data can be
done in several ways. FDA considers prospects of the evidence base for study and
approval of products. This includes the standard of proof, performed by regulations,
recommendations, guidelines, GLP (Good Laboratory Practice), GCP (Good Clinical Practice), under accurate and standardized research protocols. It can be used for
well-defined parameters of the compounds studied in toxicity tests and required for
risk assessment and design of experiments [18]. But there are some factors that cannot be clearly defined. For example, there is no sufficient information on nature of
the liver damage which occurs at the hepatic toxicity. A priori it is not clear which
factors should be taken into account for the dose modeling of this damage. From a
pathology point of view it may be necrosis, fibrosis, inflammation, etc. Which of
these data should be used for QSAR modeling? It is necessary to check carefully
the data sources, how the data were obtained and to use methods for ranking the
data quality before applying it for the toxicity modeling. Therefore, the model built
for highly specialized mechanisms or clinical measurements (e.g., the prediction of
transaminases or bilirubin increase in the blood plasma) can be more accurate and
useful. However, in the computational prediction of toxicity, these cases are rare
and it is usually required to predict more uncertain parameters. Other limitations
of modern QSAR methods are difficult to build models for organometallic compounds, complex mixtures (e.g., plant extracts), and macromolecular compounds
as polymers.
Another problem is to assess the safety of polypharmacological compounds acting on multiple targets. Also, there are significant limitations for the models building of carcinogenicity in rodents due to many existing mechanisms which may be
caused by this effect. More important is how to interpret the data obtained in the
carcinogenicity testing of drug compounds for rodents to humans. Some animal tumors have no analogs in humans [80]. If a molecule acts as a prooncogen, then it is
difficult to estimate an activity dose and tumor tissue. It is considered that creation
of the model describing toxicity or carcinogenicity based on various mechanisms
11 Computational Toxicology in Drug Discovery: Opportunities and Limitations
11.5 A Critical Assessment of Computational Approaches
for Toxicity Prediction
For a better understanding of the possible toxicity prediction it is necessary to pay
attention to a number of limitations that may occur using modern computer tools
for creation of (Q)SAR models. The major limitation factors are the guarantee of
high-quality experimental data used for creation of the training sets and understanding of what is exactly modeled by the user. When mistakes occur (e.g., an incorrect
structure of the molecule or incorrect data from toxicological studies) in the training
set, it leads to the wrong model, which provides incorrect predictions. Therefore,
considerable efforts should be made for the appropriate high-quality selection of experimental data, which will be used for creation of the model. Some recommendations for data curation in cheminformatics and QSAR modeling were published by
Fourches and co-authors [81]. Since the concept of the “most suitable” and “quality” is subjective, even among experts, determination of the quality data can be
done in several ways. FDA considers prospects of the evidence base for study and
approval of products. This includes the standard of proof, performed by regulations,
recommendations, guidelines, GLP (Good Laboratory Practice), GCP (Good Clinical Practice), under accurate and standardized research protocols. It can be used for
well-defined parameters of the compounds studied in toxicity tests and required for
risk assessment and design of experiments [18]. But there are some factors that cannot be clearly defined. For example, there is no sufficient information on nature of
the liver damage which occurs at the hepatic toxicity. A priori it is not clear which
factors should be taken into account for the dose modeling of this damage. From a
pathology point of view it may be necrosis, fibrosis, inflammation, etc. Which of
these data should be used for QSAR modeling? It is necessary to check carefully
the data sources, how the data were obtained and to use methods for ranking the
data quality before applying it for the toxicity modeling. Therefore, the model built
for highly specialized mechanisms or clinical measurements (e.g., the prediction of
transaminases or bilirubin increase in the blood plasma) can be more accurate and
useful. However, in the computational prediction of toxicity, these cases are rare
and it is usually required to predict more uncertain parameters. Other limitations
of modern QSAR methods are difficult to build models for organometallic compounds, complex mixtures (e.g., plant extracts), and macromolecular compounds
as polymers.
Another problem is to assess the safety of polypharmacological compounds acting on multiple targets. Also, there are significant limitations for the models building of carcinogenicity in rodents due to many existing mechanisms which may be
caused by this effect. More important is how to interpret the data obtained in the
carcinogenicity testing of drug compounds for rodents to humans. Some animal tumors have no analogs in humans [80]. If a molecule acts as a prooncogen, then it is
difficult to estimate an activity dose and tumor tissue. It is considered that creation
of the model describing toxicity or carcinogenicity based on various mechanisms