Computational Biology Program, NYU
December 13, 2011
Incorporating Prior Knowledge Into Network Inference
The inference of global regulatory networks is a key problem in systems biology. It has been shown that integrative approaches to biological network inference have much greater coverage and accuracy when compared to methods that are designed to work with single data-type experimental designs. The current network inference pipeline developed in the Bonneau lab, the Inferelator, already makes use of a variety of data-types. Recently, we have started to extend the Inferelator to consider prior information on regulatory interactions during the inference process. For such a procedure to be useful it must have the following two properties: 1) it must tolerate incorrect and unfit prior information, and 2) the use of prior information should not over-constrain the learning process or decrease performance on the unknown (no priors) part of the network. First results on simulated data as well as real biological data from E. coli have been promising on these critical points. Our current method shows a significant improvement in performance while being robust to false prior information.