Given the movement towards the use of alternatives to animal studies, there is a lot of new interest in in silico toxicological methods. While the ins and outs of in silico toxicology may be familiar to our typical readership, I would like to take a step back and provide foundational content to readers who may be new to in silico methods, or those who would like a refresher. Seasoned users of in silico models may also want to read on, as the acute rat oral model is discussed; and unlike the single (Quantitative) structure activity relationship (QSAR) approach for predicting mutagenicity for example, this model uses categorical QSARs to derive a consensus call to predict GHS categories.

What is it?

At the very basis of in silico toxicology is the chemical structure. This chemical structure is one key used to unlock the chemical-biological interactions that may lead to toxicity. The chemical structure is characterized by features, which define the structure. In silico toxicology uses an established link between structural features and toxicity to make a prediction. This link may be derived through a statistical association between structural features and biological activity, or through knowledge of sub-structures encoded as expert alerts (they may encode different non-reactive mechanisms). For example, the statistical acute oral toxicity model uses statistical correlations between GHS categories and structural features (together with physical and chemical descriptors) derived from a training set of over 22,000 structures. Figure 1 shows the result of a test article which is assessed using the acute rat oral statistical model. The red highlighting indicates structural features which are associated with toxicity at the GHS category queried whereas, the blue shading indicates a negative association. The model predicts that the test article is assigned to GHS category IV as this is the most severe queried category which returned a positive outcome. This positive outcome is based on the predicted value (calculate using the presence and absence of chemical features in the test compounds as well as whole molecule descriptors).

Figure 1: Assignment of GHS categories using a categorical approach.

The acute oral expert alerts use a reference set of over 40, 000 structures to support the prediction of GHS categories. In the same test article, two alerts were identified, Figure 2. These alerts map to over 500 structures with experimentally derived exact or higher GHS categories with a precision greater than 0.9. The alerts assign that test article to GHS category of V. A consensus approach is then used to derive the overall GHS category from the statistical model and the alert assignment as GHS category IV.

Figure 2: Sub-structures highlighted in red indicate those features which are identified as alerting fragments for acute oral toxicity at the GHS categories specified.

A consensus approach is then used to derive the overall GHS category from the statistical model and the alert assignment as GHS category IV.

What it is not?

In silico methods are not testing methods, and therefore, no animals are used to derive the GHS categories. In silico methods are also not black boxes! It is important to understand the structural and ideally mechanistic basis of the prediction to gain confidence in the result. The process by which we seek to interrogate the quality of the prediction is known as an expert review. There are detailed processes for performing an expert review, and these have been discussed in various publications1,2 and prior blogs3,4. Very briefly, an assessment to training/reference examples and structural analogs are important aspects. The structural similarity between the test article and an analog is considered to determine whether they are expected to exert similar toxicities. Additionally, any mechanistic information that supports the prediction would be useful. In the case of the test article assessed above, the availability of structurally similar analogs which the model predicts similarly to the test article and information included with the prediction such as the pharmacological action, histopathological findings, clinical chemistry, and in vivo findings (e.g. somnolence) for chemicals matching the alert could be used to support the reliability of the prediction.  

Figure 3: Structural analogs and the model predictions for these analogs

Performance

Bercu et al.1 assessed the fitness for purpose of the acute oral toxicity (Q)SAR models for GHS classification and labelling. In this study, the performance of the acute oral toxicity models was assessed against over 2000 proprietary and marketed chemicals. The authors found that 95% of these chemicals were predicted to a correct or more conservative category when benchmarked against experimentally derived GHS categories. The authors also detail an acute oral toxicity specific expert review and workflow to incorporate in silico models to support GHS assignment. The robustness of these models support their use in the replacement of animal studies and as contributors to the 3Rs initiative.

  1. Bercu J, Masuda-Herrera MJ, Trejo-Martin A, Hasselgren C, Lord J, Graham J, Schmitz M, Milchak L, Owens C, Lal SH, Robinson RM, Whalley S, Bellion P, Vuorinen A, Gromek K, Hawkins WA, van de Gevel I, Vriens K, Kemper R, Naven R, Ferrer P, Myatt GJ. A cross-industry collaboration to assess if acute oral toxicity (Q)SAR models are fit-for-purpose for GHS classification and labelling. Regul Toxicol Pharmacol. 2021 Mar;120:104843. doi: 10.1016/j.yrtph.2020.104843. Epub 2020 Dec 17. Erratum in: Regul Toxicol Pharmacol. 2022 Jun;131:105165. PMID: 33340644; PMCID: PMC8005249.
  • Myatt GJ, Ahlberg E, Akahori Y, Allen D, Amberg A, Anger LT, Aptula A, Auerbach S, Beilke L, Bellion P, Benigni R, Bercu J, Booth ED, Bower D, Brigo A, Burden N, Cammerer Z, Cronin MTD, Cross KP, Custer L, Dettwiler M, Dobo K, Ford KA, Fortin MC, Gad-McDonald SE, Gellatly N, Gervais V, Glover KP, Glowienke S, Van Gompel J, Gutsell S, Hardy B, Harvey JS, Hillegass J, Honma M, Hsieh JH, Hsu CW, Hughes K, Johnson C, Jolly R, Jones D, Kemper R, Kenyon MO, Kim MT, Kruhlak NL, Kulkarni SA, Kümmerer K, Leavitt P, Majer B, Masten S, Miller S, Moser J, Mumtaz M, Muster W, Neilson L, Oprea TI, Patlewicz G, Paulino A, Lo Piparo E, Powley M, Quigley DP, Reddy MV, Richarz AN, Ruiz P, Schilter B, Serafimova R, Simpson W, Stavitskaya L, Stidl R, Suarez-Rodriguez D, Szabo DT, Teasdale A, Trejo-Martin A, Valentin JP, Vuorinen A, Wall BA, Watts P, White AT, Wichard J, Witt KL, Woolley A, Woolley D, Zwickl C, Hasselgren C. In silico toxicology protocols. Regul Toxicol Pharmacol. 2018 Jul;96:1-17. doi: 10.1016/j.yrtph.2018.04.014. Epub 2018 Apr 17. PMID: 29678766; PMCID: PMC6026539.