In silico toxicology: From structure-activity relationships towards deep learning and adverse outcome pathways

Abstract

In silico toxicology is an emerging field. It gains increasing importance as research is aiming to decrease the use of animal experiments as suggested in the 3R principles by Russell and Burch. In silico toxicology is a means to identify hazards of compounds before synthesis, and thus in very early stages of drug development. For chemical industries, as well as regulatory agencies it can aid in gap-filling and guide risk minimization strategies. Techniques such as structural alerts, read-across, quantitative structure-activity relationship, machine learning, and deep learning allow to use in silico toxicology in many cases, some even when data is scarce. Especially the concept of adverse outcome pathways puts all techniques into a broader context and can elucidate predictions by mechanistic insights. This article is categorized under:Structure and Mechanism > Computational Biochemistry and BiophysicsData Science > Chemoinformatics.

Keywords: adverse outcome pathway; computational toxicology; in silico toxicology; machine learning; read across.

Figure 1

Scheme of a neural network with possible inputs. The order of the input from up to down, is in concordance with the bias introduced by the user

Figure 2

The chemical space of a QSAR model. The grey dots represent compounds from a hypothetical training dataset. The colored dots represent compounds which should be predicted by the model. The green dots represent compounds that are within the applicability domain. The pink dots represent compounds that would be predicted as out of domain as they lie on the borders of the chemical space which is very little populated. For the orange compounds, it is not clear from visual inspection how confident the model is in the extrapolation of such compounds as the surrounding area is populated, the compound, however, lies in a gap in the models' chemical space. Different applicability domain calculation might give different results here

Figure 3

Schematic drawing of an adverse outcome pathway. The pathway consists of a molecular initiating event (MIE), several key events (KE), and an adverse outcome (AO). The light grey area depicts a scheme for a linear, single adverse outcome pathway. The dark grey area depicts an adverse outcome pathways network with multiple initiating events and multiple adverse outcomes. The green square with exposure highlights that this information is not yet part of the AOP framework, however, it is highly interlinked

Figure 4

Schematic drawing of possible ways to train a machine learning model. The left path highlights that cross‐validation is possible without splitting the dataset; however, the obtained performance is only an estimate. The left path highlights that a train and test split is needed for example, if hyper‐parameters need to be tuned (in this case second inner cross‐validation can be used) or, in general, if a “real world” situation should be mimicked. It is important to note that, once the test set was used to evaluate the model performance, one should never go back to adjusting parameters with the training set to obtain a better performance on the test set as this would introduce bias into the model