Applicability domains of neural networks for toxicity prediction

Abstract

In this paper, the term “applicability domain” refers to the range of chemical compounds for which the statistical quantitative structure-activity relationship (QSAR) model can accurately predict their toxicity. This is a crucial concept in the development and practical use of these models. First, a multidisciplinary review is provided regarding the theory and practice of applicability domains in the context of toxicity problems using the classical QSAR model. Then, the advantages and improved performance of neural networks (NNs), which are the most promising machine learning algorithms, are reviewed. Within the domain of medicinal chemistry, nine different methods using NNs for toxicity prediction were compared utilizing 29 alternative artificial intelligence (AI) techniques. Similarly, seven NN-based toxicity prediction methodologies were compared to six other AI techniques within the realm of food safety, 11 NN-based methodologies were compared to 16 different AI approaches in the environmental sciences category and four specific NN-based toxicity prediction methodologies were compared to nine alternative AI techniques in the field of industrial hygiene. Within the reviewed approaches, given known toxic compound descriptors and behaviors, we observed a difficulty in being able to extrapolate and predict the effects with untested chemical compounds. Different methods can be used for unsupervised clustering, such as distance-based approaches and consensus-based decision methods. Additionally, the importance of model validation has been highlighted within a regulatory context according to the Organization for Economic Co-operation and Development (OECD) principles, to predict the toxicity of potential new drugs in medicinal chemistry, to determine the limits of detection for harmful substances in food to predict the toxicity limits of chemicals in the environment, and to predict the exposure limits to harmful substances in the workplace. Despite its importance, a thorough application of toxicity models is still restricted in the field of medicinal chemistry and is virtually overlooked in other scientific domains. Consequently, only a small proportion of the toxicity studies conducted in medicinal chemistry consider the applicability domain in their mathematical models, thereby limiting their predictive power to untested drugs. Conversely, the applicability of these models is crucial; however, this has not been sufficiently assessed in toxicity prediction or in other related areas such as food science, environmental science, and industrial hygiene. Thus, this review sheds light on the prevalent use of Neural Networks in toxicity prediction, thereby serving as a valuable resource for researchers and practitioners across these multifaceted domains that could be extended to other fields in future research.