Abstract: 

Increasing advanced analytics capabilities is a priority within big pharma, via curation of historical datasets, investment in computational infrastructure, and hiring of talent. How can these novel resources be used to meaningfully benefit patients? At Genentech we have mostly small (tens to thousands of subjects), heterogenous, highly ascertained datasets that are oftentimes used with the goal of deriving meaningful insights that can be applied to a broad population. This misalignment between the data available, the questions being asked, and the most in demand techniques (deep learning) creates unique challenges and opportunities within pharma.

This talk will demonstrate how discovery of novel biomarkers and endpoints could impact clinical development plans and why they are needed, specifically in Chronic Obstructive Pulmonary Disease (COPD). Traditional texture features derived from respiratory imaging data and clinical covariates were evaluated using standard linear models in an “added value” framework to create a benchmark model. Next, deep learning methods were attempted to try and outperform this benchmark. I will address particular concerns in implementing a prognostic deep learning model for patient stratification or enrichment such as bias, reproducibility, interpretability, and robustness. This example illustrates how high the bar is for advanced analytics to outperform traditional methods in the context of impacting a clinical development plan.

Jennifer Tom joined Genentech in 2014 as a research statistician and moved into product development as a study statistician in 2018. She has supported human genetics, microbiome efforts, imaging, early clinical development, and biomarker activities across various non-oncology indications. Prior she worked as a software engineer at Agilent and the visiting assistant Neyman Professor of statistics at Berkeley.

Bio: 

Jennifer Tom joined Genentech in 2014 as a research statistician in bioinformatics and computational biology and moved into product development as a study statistician in 2018. She has supported human genetics, microbiome efforts, imaging, early clinical development, and biomarker activities across various non-oncology indications. Prior she worked as a software engineer at Agilent and the visiting assistant Neyman Professor of statistics at Berkeley. Jennifer received a BA in Neurobiology from UC Berkeley and an MS and PhD in Biostatistics from UCLA.