Abstract

Unlike traditional talks in which the speaker presents a reasonably finished body of work, in this talk I will touch only briefly upon recently completed research, and then suggest several problems that deserve the attention of quantitatively minded researchers. The title refers to the fact that an algorithm recently developed by us for predicting the response of cancer patients to specific therapies has a lot of similarities with recent research in compressed sensing based on l_1-norm minimization. However, the proof techniques used to analyze the behavior of compressed sensing algorithms do not apply to our algorithm. Nevertheless our algorithm has performed remarkably well on a range of data sets. So the question is: Why? Another theme is that contemporary biology data sets consist of a mixture of numerical data and class labels. So new methods are needed to “integrate” all of these diverse types of data and draw meaningful conclusions. Some possible approaches to data integration will also be suggested.