Abstract

Microscopy is a cornerstone of many fields of biology. Digital imaging has increased the rate at which data are acquired, but analysis is still predominantly by visual inspection, and the images do not live on outside their original experiment or lab. We would like the corpus of existing images to be a source of large-scale analysis and mining, and to bring to other fields of biology the successes that data-driven research has brought to genetics. This requires a database of the measurements and observations---not in terms of pixels, but in terms of biological concepts: What is this object? How big is it? How are proteins distributed in it? How does it interact with other objects? How do its changes correlate? Classification and measurement, whether manual or automatic, are error-prone and inexact because the most interesting research stretches both the imaging techniques and our understanding of the biological systems. We must therefore be able to work with uncertain information, and this presents a host of new challenges: Image analysis techniques must produce probabilistic values. Data mining and machine learning techniques must work with these results, and will produce probabilistic values of their own. And the database must be able to store, index, and query these probabilistic values. As an in-depth example, I present APLA, an adaptive index structure for arbitrary probability distributions. By computing piecewise-linear approximations for the distributions, APLA can answer range queries much faster than existing solutions. It is also the first technique to efficiently answer k-NN queries on such distributions.