MACHINE LEARNING IN ‘TINY DATA’ BIOLOGY

Dr. Belinda Akpa

Assistant Professor

Integrated Synthetic and Systems Biology

North Carolina State University

The challenge of determining model parameters is a major hurdle in the development of biological models.  Parameters can only be accurately estimated with adequate data, and the amount of data required grows with the number of model parameters.  Models that attempt to bridge the length-scales of biological hierarchy (gene-protein-cell-tissue-organism) are particularly challenging to parameterize, as the complexity both within and across scales generates an explosion of model terms.  Direct measurement of the relevant quantities is likewise often difficult to achieve. However, many critical questions in biology require that we connect dynamic molecular interactions to their emergent physiological outcomes and do so quantitatively.

In my group, we frequently wrangle with models whose complexity outstrips the available data.  We model with the intention of proposing novel hypotheses and driving targeted experimental strategies to hasten discovery of causal mechanisms.  This means that we typically enter collaborative research efforts at a stage where there is little quantitative data, and further data collection may be hampered by limited resources, ethical constraints, or simply a lack of clarity as to which measurements are most likely to shed light on mechanisms of interest.  To identify plausible model parameters in these cases, we have found machine learning approaches to be of considerable value. 

Machine learning is not just a ‘big data’ concept; rather, the field offers methods that can be used to explore valid quantitative relationships even in complex systems where data is qualitative or extremely limited. In this seminar, I will highlight two case studies in which we have developed algorithms to parameterize multi-scale models from qualitative and heterogeneous data.  The first explores a critical dynamic function in plants: stoma opening as mediated by regulated vacuole fusion.  The second addresses a potential role of signaling pathway crosstalk in determining phenotype plasticity in intestinal stem cells.  In both cases, our qualitative-to-quantitative modeling approaches have provided testable, mechanistic hypotheses that have inspired new experimental strategies. 

Dr Belinda S Akpa is an Assistant Professor of Integrated Synthetic and Systems Biology at North Carolina State University.  She holds a BA, MEng, and doctorate in Chemical Engineering from the University of Cambridge (UK).  A highly interdisciplinary researcher, her current interest is in developing mathematical frameworks that integrate scarce and heterogeneous data to connect molecular phenomena to dynamic physiological outcomes. Dr. Akpa is broadly interested in mathematical biology, but more specifically in how mechanistic mathematical models can be used to inform targeted experimental strategies. By necessity, these efforts explore the limits of what one can learn from empirical observations and mathematical models, both independently and in integrative studies.