What characteristics should a QSAR model have from a regulatory point of view? A case of the ICH regulation.
As we saw in our previous post about what QSAR models are and how are they created, the REACH (chemical products) and ICH (drug) legislations not only accept, but also promote the use of alternative methods to animal testing for the evaluation of physicochemical and toxicological properties. Among these alternative methods, as you already know, are computational methods.
In this post we are going to explain the requirements that these legislations have for computational models and, more specifically, the (Q)SAR methods.
On the one hand, ECHA (European Chemicals Agency) points out on its website that registrants should avoid animal testing, and instead use “relevant and available” data sources. Among them, the REACH legislation highlights the computational methods accepted by ECHA such as the (Q)SAR methods, as we explained in a previous post.
According to this legislation, QSAR methods must follow the following 5 OECD requirements to be accepted from a regulatory point of view. They must be associated with:
- A defined parameter, e.g., mutagenicity, solubility, etc.
- An unambiguous algorithm.
- A defined applicability domain, that is, it must be clear to which molecules it can be applied.
- Appropriate performance measures (the statistical goodness-of-fit, robustness and predictivity).
- A mechanistic interpretation of human health and ecotoxicology parameters, if possible. That is, only if deemed possible, the models should give an idea of the mechanisms that give rise to the possible human health or ecotoxicology problems (or their absence) in the analyzed molecules.
At ProtoQSAR, all our models fulfill these 5 quality parameters. This is a very important point, since using models that do not comply with these requirements would surely lead to the rejection of registration reports sent to regulatory agencies. In addition, we ensure as much as possible that the methods that have been used to obtain the experimental data used to generate the models follow the experimental protocols specified in the OECD guidelines. If you wish to know more about these guidelines, a link to them is provided in the references.
The ICH standard also includes (Q)SAR methods within the intended workflow to assess the mutagenicity and carcinogenic potential of drug impurities.
The following image outlines how the ICH regulations use (Q)SAR methods for classifying impurities.
- If there is in vitro or in vivo data according to which a substance is mutagenic, it is classified as “class 1” (if it is carcinogenic) or “class 2” (if it is not or this information is unknown).
- If no experimental data exist, an in silico prediction should be made. Based on these findings, the substance will be classified as “class 3” or “class 4”.
- In the event that the experimental or in silico methods give the molecule as neither mutagenic nor carcinogenic, it will be classified as “class 5”.
Each of these classes according to the ICH-M7 regulation has subsequent requirements for its final acceptance. Classes 4 and 5 are considered non-mutagenic and therefore are regulatory acceptable.
As seen in the figure, we may have three situations, according to the results of these in silico studies:
- Likely to conclude positive
- Likely to conclude negative
- Uncertain result
How do we arrive at one of these three conclusions from our models? What the regulations require from us is to apply two types of models: SAR models (rule-based) and QSAR models (statistical). Hence, the same regulations group them as (Q)SAR studies.
In the following image we explain what we may conclude from these two types of tests:
If the two types of models agree, we will conclude that the substance is “likely positive” or “likely negative”.
It may also happen that the prediction of the SAR method is positive while the result of QSAR method is not conclusive (result “out of domain”), in which case we will conclude that the substance is “likely positive” according to the result of the SAR method.
In the rest of the cases, we have no option but to indicate that the result is uncertain and, therefore, it will be necessary to carry out in vitro or in vivo experiments, as the last option.
 ECHA. “Alternatives to animal testing under REACH” Retrieved October 25, 2021a (https://echa.europa.eu/animal-testing-under-reach).
 OECD. “OECD Test Guidelines for Chemicals – OECD.” Retrieved October 25, 2021b (https://www.oecd.org/chemicalsafety/testing/oecdguidelinesforthetestingofchemicals.htm).
 Barber, Chris, Alexander Amberg, Laura Custer, Krista L. Dobo, Susanne Glowienke, Jacky Van Gompel, Steve Gutsell, Jim Harvey, Masamitsu Honma, Michelle O. Kenyon, Naomi Kruhlak, Wolfgang Muster, Lidiya Stavitskaya, Andrew Teasdale, Jonathan Vessey, and Joerg Wichard. 2015. “Establishing Best Practise in the Application of Expert Review of Mutagenicity under ICH M7.” Regulatory Toxicology and Pharmacology 73(1):367–77. doi: 10.1016/J.YRTPH.2015.07.018.
 Macmillan, Donna S., Steven J. Canipa, Martyn L. Chilton, Richard V. Williams, and Christopher G. Barber. 2016. “Predicting Skin Sensitisation Using a Decision Tree Integrated Testing Strategy with an in Silico Model and in Chemico/in Vitro Assays.” Regulatory Toxicology and Pharmacology 76:30–38. doi: 10.1016/J.YRTPH.2016.01.009.