Interpretable machine learning for biological discovery

Susan Liao, PhD

NYU Courant Institute of Mathematical Sciences, Postdoctoral Fellow with Jef Boeke, PhD, and Additional Ventures LSRF Fellow

Abstract: Neural networks hold great promise for deciphering complex biological logic due to their expressive nature that can describe a breadth of relationships in data that can be quantitatively described through mathematical functions. However, despite great success with predicting outcomes, most neural networks fail to advance biological discovery as they were designed with primary goal of prediction accuracy. Despite excellent predictions, the vast majority of neural networks cannot explain how they arrived these predictions. It is difficult to trust or verify whether these predictions reflect true biological relationships or instead capture biases in the dataset. Uninterpretable neural networks thus provide limited mechanistic insight into underlying biological processes. Furthermore, predictions from an uninterpretable network may not generalize, failing to predict outcomes from data that it has not seen before, limiting the utility of the network in basic research or clinical settings. There is a clear need to design interpretable neural networks that will enable biologists to generate and test novel hypotheses from their data. In this presentation I will present our progress towards this goal, focusing on how advances in data acquisition and data analysis can improve neural network interpretability. These innovations will address interpretability concerns across a wide range of biological questions involving gene expression in both development and disease.

Roles of negative and positive feedback in cardiovascular adaptations

Linda Irons, PhD

Department of Biomedical Engineering, Yale University, Postdoctoral fellow with Jay Humphrey, PhD

Abstract: Single ventricle defects result in altered hemodynamics and reduced blood oxygenation, both of which can stimulate remodeling responses throughout the systemic and pulmonary circulations. Such changes, in turn, can adversely affect the delivery of oxygenated blood within end organs at appropriate pressures and flows, hence contributing to progressive morbidity and mortality. There is, therefore, a pressing need to understand not just how the hemodynamics differs in single ventricle disease, but also how these alterations drive widespread remodeling that is often maladaptive. Toward this end, we have developed computational models “from transcript to tissue” to simulate how cell-perceived stimuli result in particular transcriptional changes that in turn drive tissue level changes in structure and function. Whereas homeostatic processes drive adaptive remodeling through negative feedback, compromised or lost hemostasis results in maladaptive responses characterized by positive feedback.