The methods called quantitative structure-activity relationship (QSAR) are based on the assumption that the activity of a certain chemical compound is related to its structure. More precisely, this approach says that the activity, or the property, for instance the toxic effect, is related to the chemical structure through a certain mathematical algorithm, or rule. QSAR models are also called in silico methods, which actually refer to a somehow broader set of methods. Sometimes the expression SAR is used, which refers to methods which are non quantitative, but only structure-activity relationship. For instance, this applies to models where the toxicity, not expressed in a quantitative way, but only as category (such as toxic or not toxic) is linked to the occurrence of a given fragment in the molecule.
Results of the QSAR models depend on the property to be evaluated, on the individual QSAR model, on the chemical to be evaluated and in some cases on the user, for the models where parameters have to be fixed by the user, or decisions have to be taken on the result validity. The general evaluation of a model is typically done using a set of chemicals, using statistical methods, such as internal and external validation. This will provide an evaluation on the general performance of the model. In addition, a careful evaluation on the specific use for the target chemical should be done. If the user has no experience, it is advisable to ask advice.
Read-across is a non-testing method, like QSAR models. In case of read-across the evaluation is simply done considering one or few chemicals, similar to the target compound. The assumption is that the property of the target chemical to be evaluated is similar to the properties of the similar compound(s).
REACH is the legislation that regulate the market of chemical substances, produced and imported. All chemical substances require characterization of their environmental and toxicological properties, before their use in Europe. This is a huge effort which requires chemical industry to notify the properties of their chemicals.
REACH, since its first article, promotes innovation, and mentions that new, alternative methods should be searched to improve the evaluation of the chemicals. QSAR methods are included in the alternative methods. Annex XI of REACH defines the conditions to use QSAR models. Within these conditions, the use should evaluate the following points: The QSAR model should be scientifically valid. The evaluation has to be done on the use of the model for a defined chemical (check of the applicability domain). Indeed, it may be that a model is appropriate for a chemical, but not for another. The model should be suitable for risk assessment or classification and labeling. Thus, the model has to adhere to the requirement of REACH for the particular data which are defined by the regulation. It is important to provide suitable documentation. Thus, simply copying the value obtained by a certain model without proper evaluation of the model results may be not sufficient.
REACH regulates the chemical substances. REACH and ECHA (the European Chemicals Agency) do no provide a list of suitable QSAR models. Within Annex XI REACH clearly states that appropriate evaluation of the applicability domain has to be done. Indeed, it may be that a model is appropriate for a chemical, but not for another. Thus, the issue is more complex than simply listing good or bad models. The reliability of the QSAR models strongly depends on the property. In case of physico-chemical properties, mainly in cases where there are thousands of data, good models exist. For this kind of endpoint, the uncertainty of the experimental data is limited. In case of more complex endpoints, and in particular for chronic endpoints for human toxicity, the number of data is more limited, and the uncertainty of the experimental values is much higher. ANTARES will evaluate QSAR models for REACH, check them, and will provide a list of tens of QSAR models which can be used, provided that the chemical of interest is within the applicability domain of the model.