SGER: Developing Learning Theory for Genetic Network Inference

Columbia University, New York NY

Investigators

Abstract

The recent explosion in the number of papers published on practical and theoretical aspects of inference from data and data mining attests to the great progress in our understanding of learning theory. However, some glaring holes in understanding still exist, and, while the number of fields of application continues to rise, literature addressing conceptual problems is scant. The most troubling gap is that there is, in fact, no unique theory of learning yet developed: the number of principled approaches is nearing a dozen. Moreover, purely empirical, highly specialized techniques are proliferating innumerably. A few attempts have been made to build a common foundation for all methods, but a full under-standing is still lacking. Similarly, even though equivalence of many approaches is presumed, it is clear that for purely practical reasons some techniques may be superior to the others in different settings. In complex data analysis tasks, it might be beneficial to use different approaches for different sub-tasks. However, this rarely happens: most authors' expertise is highly specialized, and without a unifying theory there is a fear of less familiar methods and a general unwillingness to use them together with the more common ones. On the other hand, there exists a broad field of problems for which a uniform understanding and simultaneous usage of many methods is essential. Currently, one of the most pressing such tasks is learning genetic regulatory pathways from microarray experiments. These experiments are characterized by extreme undersampling; thus learning methods with inherent capacity controls should be used. Straightforward in-corporation of prior biological knowledge is also absolutely essential. Further, astronomical amounts of data supplied by the experiments require methods that are either analytically simple or computationally efficient. Both traits, however, are hampered by non-uniformities in the data (e.g., missing or irregular samples). None of the current methods alone is capable of addressing these disparate issues. Only a careful synthesis, grounded in a good understanding of common foundations and practical differences, can be successful. The proposed research is designed to address all of the above-mentioned fundamental issues in learning theory, with the inference of genetic regulatory networks as a useful testbed.