Myrtle Davis, scientific vice president of discovery toxicology at Bristol Myers Squibb (BMS), has had an impressive 30-year toxicology career. After a postdoctoral position in toxicologic pathology, she became a faculty member at the University of Maryland. Before joining BMS, she was a research toxicologist at Eli Lilly and later became the chief of the toxicology pharmacology branch of the developmental therapeutics program at the National Cancer Institute. Having worked in the academia, government, and industry sectors, Davis has been instrumental in leveraging her expertise to advance several drug candidates to Phase 1 clinical trials and beyond. Her commitment to enhancing patients’ experience through optimized drug development workflows remains at the forefront of her work. “We have limitations in what we can predict from non-clinical experiments,” she said. “The challenge is making sure our clinical data feeds back into the non-clinical space so we continue to improve patient outcomes.”
What does assessing the toxicology profile of a new molecule look like at BMS?
The discovery process starts with the target. Those who study diseases look at how the target contributes to disease. We look at the opposite side and ask what the target does in a healthy tissue or cell and what happens if that role gets compromised. We then test whether modulating the target in normal tissues results in toxicity. Our work continues into the Phase 3 stage of clinical development, where we examine toxicities that may occur after several months of treatment or in vitro exposure, building evidence alongside clinical data for adverse effects. For example, diacylglycerol kinase was one of our recent targets. We determined the role of this protein in normal cells and conducted experiments to see whether modulating it made those cells dysfunctional. Once we selected a molecule that appeared promising, we progressed to early drug development studies to support a Phase 1 trial, which helped us determine the dosing schedule. We also identified biomarkers to monitor adverse effects, which we implemented successfully.
What types of in vitro assays do you use in toxicity evaluations?
Once we know where the target is expressed, our next step is to identify the specific cell that expresses it. We then use an in vitro assay system to evaluate the effect of modulating the target in a physiologically relevant context. These cell culture experiments, whether 3D, 2D, or using tissue slices, are designed to mimic what we expect in vivo. We use a variety of assays, each tailored to answer specific questions and ensure physiological relevance, ranging from simple cell viability to complex pathway analyses.
How do you decide when to move from in vitro to in vivo experiments?
Moving to in vivo studies depends on whether there are questions we cannot address in vitro, and there will be several. We make this decision based on the physiology and the normal role of the target in tissues and cells. Another factor in this decision is regulatory expectations. When there is a relevant model for evaluating human safety, we have to use it to understand potential liabilities, toxicities, risks, and other concerns.
How do you assess the potential long-term effects of a new drug?
Predicting long-term outcomes based on early in vitro or other short-term systems is very challenging. However, we can use vigilant systems like the vaccine adverse effects reporting system and our diligence missions that monitor patient outcomes in the clinic. Clinical data provide the best insights into what might happen in the long term. Our focus is on using clinical outcomes from Phase 1, 2, and 3 clinical trials to predict what might occur after two or three years of treatment. We then bring this information back into early discovery to refine our models and improve our predictive capabilities.
How do you use predictive toxicology in toxicology evaluations?
The word “prediction” is a word we would love to use with confidence. For a predictive model, we're working on integrating all available data, including clinical data. We gather and consolidate this information with advanced technologies that allow us to analyze and identify potential patterns that could help us foresee future outcomes. A critical aspect of this process is model validation. To make reliable predictions, we need enough data to validate our models, both in clinical settings and earlier phases. This involves comparing predicted outcomes with actual results to ensure the models' accuracy. That's exciting because we have so much data, and the new technologies allow us to see patterns that help us make predictions.
How are you using artificial intelligence (AI) and machine learning for determining toxicity?
We build machine learning models that can learn from our data. Then we train our toxicologists to make decisions from the complex data sources that these machines deconvolute. For example, we're using AI to reduce the number of studies we need. Our chemistry group is focused on rapid in silico evaluations of molecular features from molecules. Without synthesizing anything, we can define features that influence the pharmacokinetics and toxicities that we might see in vivo. In pathology, AI helps by quickly categorizing images from slides as normal or abnormal. This speeds up diagnoses for pathologists. We digitized our histology and discovery toxicology data early on, allowing us to apply deep neural network algorithms to assist pathologists and define the limitations of these networks. Our early digitization means we have extensive data for the algorithms to learn from, which helps us progress more quickly.
What other emerging trends in toxicology do you think will affect drug development?
We're currently working with spatial transcriptomics and finding it incredibly powerful. It allows us to see the gene signature within each cell in context, which gives us an atlas of information. For instance, when we need to know where a target is expressed, we no longer need to conduct an experiment; we can simply refer to this data source. Spatial transcriptomics helps us understand where the target is and how pathways might affect other cells. This capability opens questions that we can explore in silico. Additionally, I've seen impressive applications of quantitative structure-activity relationship (QSAR) methods combined with other AI techniques. Traditional QSAR models, when integrated with these new methods, show incredible promise. We're keeping an eye on how we might use these advancements at BMS.
What has been the most rewarding part of helping characterize the toxicity profiles of new drugs?
Patient experience is our beacon. We know that no prediction is absolute and that no technology alone will solve all problems and answer all questions. When we enter a Phase 1 or 2 trial, and the dose that we selected proves to be accurate, we avoid adverse effects in the first dose. This confirms that we've chosen a promising molecule to move forward. Looking back at past drugs, we see cases where the adverse effects were significant, often not much better than the disease itself in terms of patient experience. My satisfaction comes from hearing a patient say, "Your drug has been manageable; I've been okay." Knowing that our work contributes to a better patient experience is incredibly fulfilling.
This interview has been condensed and edited for clarity.
