Chair:

Kamel Mansouri, NIEHS/NTP

Co-Chair:

Denis Fourches, Oerth Bio

Primary Endorser: Biological Modeling Specialty Section


Endorser(s): Arab Toxicologists Association Special Interest Group


Endorser(s): Computational Toxicology Specialty Section

Humans are exposed to an ever-increasing number of chemicals, but only a fraction of these chemicals have been evaluated for potential risks to human health and the environment. Thus, both regulators and manufacturers need rapid and efficient approaches to evaluate the toxicity of chemicals currently on the market as well as those in development. Recent advances in information technology and machine learning, in addition to the rapid increase of experimental data, have fostered the development of in silico approaches that leverage the relationships between chemical structures and their biological activities. However, the use of these tools still suffers from the lack of understanding of the theory behind them and misinterpretation of the results. In this session, we will focus on the importance of the concept of chemical similarity that is the basis of the most widely used in silico approaches, such as quantitative structure-activity relationships (QSARs), read-across, clustering, and classification approaches, among others. These approaches are based on the congenericity principle, which is the assumption that chemicals with similar structures (congeners/analogs) also are associated with similar biological activities. This concept is often oversimplified as the visual look-alike, which ignores the subtle differences in structural characteristics that can lead to major differences in biological activity. This is one of the reasons the same analogs are sometimes used in read-across for different endpoints based only on a similarity index calculated using a general fingerprint. Thus, the goals of this session are to (1) address the misconceptions about chemical similarity by redefining it based on use, context, and how it differs from the perspectives of medicinal chemists, model developers, and regulators; (2) demystify the machine-learning approaches to black boxes and tools relating structures to activity by identifying the most influential structural features that define similarity for a specific outcome; and (3) identify the best practices regarding how and when to apply machine-learning and class-based approaches depending on the goal and context of use. The first speaker will give an introductory presentation defining the different relevant concepts that play a role in the design and use of the similarity and class-based approaches. The second speaker will provide an overview of the different tools used to encode chemical information. The following talk will address the concept of similarity and the fundamentals of supervised and unsupervised approaches. The fourth speaker will discuss the endpoint-specific similarity measures in relation to the selection of the most adequate analogs. The fifth speaker will highlight the best practices for applying different approaches to example studies. The last speaker will present on the application of similarity measures for supervised and unsupervised approaches.