Machine Learning Approaches in Computational Toxicology Studies

In the present world, the use of computational toxicity assessment techniques is highly encouraged as an alternative to the standard experimental toxicity testing. Quantitative structure–activity (toxicity) relationships (QSA(T)R) modeling is the most widely used in silico technique for risk assessment and hazard identification of chemicals. Recent advances in the QSAR technique include the application of several advanced machine learning techniques, capable of capturing the true (linear or nonlinear) relationships or patterns. In the present chapter, the authors describe several popular machine learning techniques in a simple and illustrative manner. The chapter initiates with an introduction to QSAR and machine learning techniques in toxicity testing along with its brief history and recent advances. Then the chapter progresses with the discussion of basic steps that are involved in toxicity data set collection, preparation (curation), and calculation of descriptors and fingerprints (or features) representing the collected chemicals. Subsequently, several often-used machine learning techniques under three broad categories, namely, unsupervised, supervised, and semi-supervised learning are discussed. In the next segment, the appropriate model selection and validation approaches applied to a QSAR model are briefly discussed. Finally, the freely available software tools and open-source libraries that are relevant to machine learning are highlighted.