Are QSAR models available for all endpoints?

in

QSAR models are available and in use for some endpoints, while being limited for others (e.g. long term mammalian toxicity).  There are several reasons, including:

  1. the limited amount of high quality experimental data currently available (and released by industry) for some endpoints,
  2. the higher complexity of the chemical and biological processes involved in generating the toxicity for some endpoints, and
  3. the lower regulatory acceptance for certain endpoints.  

In general, for physico-chemical properties QSAR models are more reliable, because the phenomenon to be modelled is less complex, and thousands of experimental results are available in the databases.  Whereas, for models to predict chronic effects, the data are much more limited, and the phenomenon is much more complex.

The EC funded ANTARES project is evaluating existing models which could be used for REACH.  Hundreds of possible models have been identified and listed, which theoretically could be used for tens of REACH endpoints. ANTARES is currently checking the performance of the QSAR models for several endpoints.  (See: FAQs: ‘Is there an independent review of QSAR models’ and ‘What makes a good QSAR model’.)

Related FAQ

Are QSAR models expensive, or free to use?

The CAESAR and VEGA platforms have been produced from EC-funded research, and are therefore freely available for use.  The software can be downloaded. Furthermore, predicted values for more than four million chemicals will be made freely available.

The Danish Environmental Protection Agency has created and made available a database with predictions from more than 70 (Q)SAR models on endpoints for physico-chemical properties, fate, eco-toxicity, absorption, metabolism and toxicity.  More than half of all the estimates are for mammalian (human) toxicity endpoints and include commercial data sets from TOPKAT and MULTICASE as well as models developed in-house.  A structure set of about 166.000 discrete organic chemicals, including almost half of the (around) 100,000 EINECS substances, has been batched through the models, and the results are integrated in the database.  (EINECS: European INventory of Existing Commercial chemical Substances.)

The US Environmental Protection Agency made available a series of QSAR models, within the EPISuite and the T.E.S.T. platforms.

The EC funded ANTARES project is evaluating existing models which could be used for REACH, many of which are free to use.  Hundreds of possible models have been identified and listed, which theoretically could be used for tens of REACH endpoints.

The EU Joint Research Centre has promoted the availability of reliable computer-based estimation methods for use in the regulatory assessment of chemicals.  Within the JRC Institute for Health and Consumer Protection (IHCP), the European Chemicals Bureau (ECB) developed a range of freely available software.  One of the aims of the IHCP is to support the implementation of EU chemicals policy (including the safety assessment of industrial chemicals, chemicals in consumer products, pesticides and biocides) through the development, assessment and application of computational (in silico) methods. 

Typically in such QSAR models, the training set used to develop the model, the algorithm, and other aspects of the model are open for scrutiny by the regulator.   However, in commercially available software the full documentation of each model may not be available.  Indeed, the algorithm and the training set used to build the model are often confidential.  Despite this, the use of commercial programs has not been explicitly criticised by regulatory authorities.  In reality, commercial programs have been used in the USA and Europe, because of their ease of use, their availability and because they represent a major source of QSAR models.  We do not foresee restrictions in their use, and in our opinion it would be a pity to renounce these models.  However, it is likely that in the case of two similar models being available, one commercial and one freely available, the user will prefer the second one.  Within REACH, it also seems likely that the regulators will demand more transparency from the commercial models, and/or prefer the use of fully documented models.

What makes a good QSAR model?

According to REACH regulation (Annex XI) an assessment using a QSAR model is valid if:

  1. the model is recognised as scientifically valid;
  2. the substance is included in the applicability domain of the model;
  3. results are adequate for classification and labelling and for risk assessment;
  4. adequate documentation of the methods is provided.’

These four ‘conditions’ are specific to the regulatory context, and address important scientific and practical aspects of the use of in silico models.  It is important to note that the four conditions are not just about the model, but also about the particular chemicals it is used for in an assessment, the particular regulatory function it is used for, and how well the model and its use are documented.  

A QSAR model is built on a set of experimental data for particular chemicals (a ‘training set’) and it is trialled by testing other chemicals for which there is also experimental data available (the ‘test set’).  In this way a QSAR model is designed to evaluate a particular set of chemicals in relation to a particular endpoint.  So while it could be used for other chemicals outside this ‘applicability domain’, the evaluation of those chemicals will have a higher level of uncertainty.  

For this reason, REACH and ECHA do not provide approval for models: each use of QSAR models has to be evaluated on its own merit, on a case-by-case basis.  

The first of the four conditions above is nevertheless about the model itself.  In the implementation of REACH, ECHA and others usefully refer to the five OECD principles for QSAR models.  The OECD stated that ‘to facilitate the consideration of a (Q)SAR model for regulatory purposes, it should be associated with the following information:

  1. a defined endpoint;
  2. an unambiguous algorithm;
  3. a defined domain of applicability;
  4. appropriate measures of goodness-of-fit, robustness and predictivity;
  5. a mechanistic interpretation, if possible.’

These five principles refer to the rigour of the model and the transparency of the information provided by the model developer.  The REACH demand to provide ‘adequate documentation of the models’ and their use, is central to demanding and ensuring the quality of models and the quality of their use, now and in the future.  It is ultimately about enabling the regulator to evaluate the chemical safety assessments independently in the interest of human health and the environment.

Is there an independent review of QSAR models?

ECHA does not intend to produce a list of ‘approved’ models, because the value of a model depends on how it is used.  (See FAQ: ‘What makes a good QSAR model?’)  Every user of QSARs needs to be aware that QSAR models are only appropriate and reliable for specific sets of chemicals.  A highly reliable model will not produce reliable results for chemicals that lie outside the domain of applicability.  In addition, models may be suitable for different regulatory functions: risk assessment, classification and labelling or prioritisation, because each makes different demands on the model.  

In terms of the form of the outputs, the QSAR Model Reporting Format (QMRF) checks that a certain number of pieces of information are given.  However, such an assessment is carried out on the basis of the values provided by the developer of the model, and these are not checked independently.

The EC funded ANTARES project is evaluating existing models which could be used for REACH. Hundreds of possible models have been identified and listed, which theoretically could be used for tens of REACH endpoints. ANTARES is currently checking the performance of the QSAR models for several endpoints: carcinogenicity, mutagenicity, LD50, fish and daphnia acture toxicity, bioconcentration factor, ready biodegradability and logP.

Given the range of endpoints, the vast numbers of chemicals, the large palette of mathematical algorithms available, and the potential to use tens of thousands of chemical fragments and thousands of chemical descriptors to build a predictive model, it is clearly possible for future scientists to generate huge numbers of QSAR models.  It is easy to imagine an explosion in the number of useable models, many with quite similar performance.  

Further information

1) ANTARES project.  www.antares-life.eu

What's next

Learn more about in silico, browse resources,
find out what's happening and get involved