2021 Annual Meeting Report: Artificial Intelligence in Toxicology


By Sushant Kamath posted 03-25-2021 12:13


“If our era is the next Industrial Revolution, as many claim, AI is surely one of its driving forces.” –Fei-Fei Li1

Artificial intelligence (AI) is surely the “driving force” in many areas of industrial application, but more so in the field of toxicology. This was more evident during one of the first Symposium Sessions at the Virtual SOT 60th Annual Meeting and ToxExpo, titled “Industrial Applications of Artificial Intelligence in Toxicology.” The session was chaired by Dr. Catrin Hasselgren, Principal Scientist at Genentech Inc., and Dr. Nigel Greene, Director, Data Science & AI, Clinical Pharmacology & Safety Sciences, at AstraZeneca. Dr. Hasselgren spoke about the evolution of AI, which is not a new concept, has existed since the 1950s, and has evolved over the years with an exponential growth in computational processing speed, resulting in the massive AI boom that we are now witnessing. The data explosion is creating the need for AI and machine learning. Nonclinical data generated from multiple sources like ’omics technologies, cell painting, imaging of tissues, digital pathology, and noninvasive in vivo monitoring all need to be compiled for submission to regulators. So how do we make sense of all this data? This explosion of data is beyond the capacity of the human brain to digest and resolve. AI is the only way forward. With this brief introduction, Dr. Hasselgren opened the session and introduced the distinguished speakers.

Dr. Alexander Amberg, Head of In Silico Toxicology, Preclinical Safety, at Sanofi, presented how his team developed in silico models from in vivo drug toxicity data and how they were successfully applied for regulatory submissions. From a regulatory context, in silico approaches play only a minor role, with the exception of potential mutagenic impurities according to ICH M7, with well-established application of in silico methods for regulatory submissions. However, when it comes to in silico methods for predicting in vivo organ toxicity, improvements and expansion in computational models are needed for regulatory acceptance. The goal is to develop internationally harmonized and regulatory-accepted in silico toxicology protocols. This project is lead by a consortium of 60 partners and funded through a National Institutes of Health grant, and currently two protocols, for genetic toxicology and skin sensitization, have been finalized. Protocols for liver, cardiac, renal, pulmonary, and neurotoxicity are under development. The models are being developed based on regulatory-approved in vivo data from sources like eTOX, PharmaPendium, and Leadscope. Dr. Amberg described in detail how these models are being developed and validated and also presented a case study for prediction of liver toxicity.

Dr. Fangyao Hu, Scientist in the Department of Safety Assessment at Genentech, presented the use of deep learning methods to enhance toxicity assessment of histopathological images. He provided a brief introduction to deep learning and described how to relate in vitro data to clinical toxicity. Dr. Hu described in detail how the methods are used to detect objects from histopathology images to evaluate “study-level” drug safety and then classify them to evaluate “slide-level” drug safety.

Dr. Marinka Zitnik, Assistant Professor of Biomedical Informatics at Harvard Medical School, spoke about actionable machine learning for safe and effective therapeutics. She emphasized the need for modern instruments for data-intensive science for knowledge discovery, which is now going virtual. It is a challenge to realize this vision because there is a need to integrate heterogenous confounded data that can span from molecules to society, from which we need to understand disease treatment mechanisms, drug safety, and efficacy. The challenge lies in making this data operational and amenable to data analytics. Dr. Zitnik’s lab is leveraging “Networks” as a general language for interconnected data. The Networks are used to connect data from cells to individuals to populations spanning multiple scales. Dr. Zitnik’s lab is pioneering “Graph Neural Networks,” which is a new class of deep learning models. She described the core technical ideas behind this approach and explained how this is being applied to obtain safer and more effective medicines and drug combinations. Further, it also is being applied in drug repurposing for COVID-19 treatment.

The final presentation of the session was from Dr. Andreas Bender, Natural Philosopher for Molecular Informatics, Department of Chemistry, University of Cambridge. He described the use of ’omics data and pharmacokinetic information in toxicity and safety prediction. His perspective of looking at the entire concept was interesting, as he described it as one where “the world is flat” and one where “the world is round.” Dr. Bender explained how the molecular structure is connected to proteins through bioactivity data and how the proteins are connected to the phenotype through pathways and the phenotype can be connected back to the molecular structure through the phenotypic response data. This is the simple view of the “flat world.” But the reality is that “the world is not flat.” In terms of data, links between drugs and diseases, although quantitative, are incompletely characterized; subtle differences in compound effects need to be addressed; phenotyping is sparse and subjective; and there are many other issues that need to be addressed. These issues can be resolved by AI. Dr. Bender explained the use of ’omics data in drug-induced vascular injury (DIVI) and in drug-induced liver injury (DILI) and how gene expression data and pharmacokinetics approximations, respectively, are providing the solutions. He also explained the different machine-learning models for pharmacokinetics.

To summarize this Symposium Session, AI is a solution to improve efficiencies in drug discovery and development. With rising costs and tighter regulations in chemical safety, as well as to reduce animal usage, it is time to embrace the new path forward.

1Li, Fei-Fei. 2018. “How to Make A.I. That’s Good for People.” New York Times. https://www.nytimes.com/2018/03/07/opinion/artificial-intelligence-human.html.

This blog was prepared by an SOT Reporter and represents the views of the author. SOT Reporters are SOT members who volunteer to write about sessions and events in which they participate during the SOT Annual Meeting and ToxExpo. SOT does not propose or endorse any position by posting this article. If you are interested in participating in the SOT Reporter program in the future, please email Giuliana Macaluso.

Sessions delivered during the 2021 SOT Annual Meeting and ToxExpo will be available on the Virtual Meeting Platform on demand to registrants through May 31, 2021.