18 Integrating QSAR, Read-Across, and Screening Tools …
375
18.3 Results
18.3.1 The VEGA Software
VEGA represents a sophisticated system to assist the user in the evaluation of the
properties of chemical substances. VEGA provides three kinds of outputs, which
conceptually are quite different: QSAR models, measurement of the reliability of
the predictions, and read-across.
As the other systems with QSAR models, VEGA provides predictions. We try
to offer models for a series of endpoints and if possible to give more than one
model for the same endpoint. The models have been developed within a series of
projects funded by the European Commission (EC): CAESAR, ANTARES, CALEIDOS, PROSIL, ORCHESTRA, ToxBank, EU-ToxRisk, COMBASE, VERMEER,
and EDESIA. Further funds have been received by the German UBA and the German
Federal Ministry for the Environment, Nature Conservation, Building and Nuclear
Safety, within the projects PROMETHEUS, JANUS and toDIVINE, and by the Italian Ministero della Salute.
We also included some well-known models from other platforms. For instance, we
have models replicated from EPISuite, or ToxTree. Some models have been developed in collaboration with US EPA, as in the case of the model for developmental
toxicity. The model for reprotoxicity has been implemented based on the software
developed by Procter and Gamble researchers [19], and we already mentioned the
implementation of the EnviPath software within JANUS.
This is to say that we try to offer to the user good models, regardless of the
theoretical assumptions and the sources, but looking at the quality of the model,
in terms of quality of the data at its basis and statistical parameters. The statistical
values of the models have been published in a series of papers [20–25].
VEGA provides an advanced measurement of the applicability domain (AD). It
has been shown that the prediction is more reliable if we take into account the ADI.
In this way, the user has a clear, automatic way to evaluate the reliability of the
prediction, and the use of the ADI increases the confidence of the prediction. The
software at the basis of the ADI is orthogonal compared to the software that makes
the prediction.
Furthermore, VEGA can be used as a simple tool for read-across. Indeed, the
user can even disregard the prediction and make his or her assessment based on the
most similar compounds. This feature is probably not well-recognized by the user,
but it is very simple. Also in this case, if there is more than one model for the same
endpoint, we recommend use all of them. Indeed, the software for the ADI refers to
the chemicals present in the different datasets at the basis of each model, and thus,
each model may show different chemicals. It may also happen that different values
are present in the different sets of chemicals used by the different models. In these
cases, we left the original experimental values as used by the developers.
Behind the list of the most similar compounds with the experimental values, there
is an additional tool for read-across which is present for the models that use SA. In this
375
18.3 Results
18.3.1 The VEGA Software
VEGA represents a sophisticated system to assist the user in the evaluation of the
properties of chemical substances. VEGA provides three kinds of outputs, which
conceptually are quite different: QSAR models, measurement of the reliability of
the predictions, and read-across.
As the other systems with QSAR models, VEGA provides predictions. We try
to offer models for a series of endpoints and if possible to give more than one
model for the same endpoint. The models have been developed within a series of
projects funded by the European Commission (EC): CAESAR, ANTARES, CALEIDOS, PROSIL, ORCHESTRA, ToxBank, EU-ToxRisk, COMBASE, VERMEER,
and EDESIA. Further funds have been received by the German UBA and the German
Federal Ministry for the Environment, Nature Conservation, Building and Nuclear
Safety, within the projects PROMETHEUS, JANUS and toDIVINE, and by the Italian Ministero della Salute.
We also included some well-known models from other platforms. For instance, we
have models replicated from EPISuite, or ToxTree. Some models have been developed in collaboration with US EPA, as in the case of the model for developmental
toxicity. The model for reprotoxicity has been implemented based on the software
developed by Procter and Gamble researchers [19], and we already mentioned the
implementation of the EnviPath software within JANUS.
This is to say that we try to offer to the user good models, regardless of the
theoretical assumptions and the sources, but looking at the quality of the model,
in terms of quality of the data at its basis and statistical parameters. The statistical
values of the models have been published in a series of papers [20–25].
VEGA provides an advanced measurement of the applicability domain (AD). It
has been shown that the prediction is more reliable if we take into account the ADI.
In this way, the user has a clear, automatic way to evaluate the reliability of the
prediction, and the use of the ADI increases the confidence of the prediction. The
software at the basis of the ADI is orthogonal compared to the software that makes
the prediction.
Furthermore, VEGA can be used as a simple tool for read-across. Indeed, the
user can even disregard the prediction and make his or her assessment based on the
most similar compounds. This feature is probably not well-recognized by the user,
but it is very simple. Also in this case, if there is more than one model for the same
endpoint, we recommend use all of them. Indeed, the software for the ADI refers to
the chemicals present in the different datasets at the basis of each model, and thus,
each model may show different chemicals. It may also happen that different values
are present in the different sets of chemicals used by the different models. In these
cases, we left the original experimental values as used by the developers.
Behind the list of the most similar compounds with the experimental values, there
is an additional tool for read-across which is present for the models that use SA. In this