Modern systems are known to be massive-scale, with a hierarchy of dynamic and unknown topologies. The behavior of these systems can be captured via dynamic graphical models (DGM). An important application of DGMs is in the inference of dynamic gene regulatory networks that may change spatially across different cells and over time in response to different physiological cues. This work aims at developing efficient computational tools for the inference of DGMs that are not only statistically sound, but also adaptive, parallelizable, and implementable in massive scales. Much of the progress in the inference of DGMs with structural constraints is based on the maximum likelihood estimation (MLE) with relaxed regularization, which neither result in ideal statistical properties nor scale to dimensions encountered in realistic settings. This work addresses these challenges by departing from the regularized MLE paradigm and resorting to a new class of constrained and combinatorial optimization that can systematically learn DGMs in near-linear time and memory. For the first time, we can infer instances of DGM with more than 500M variables in less than an hour. We will also show the promise of our method in the context of inferring gene networks underlying oncogenesis, using Glioblastoma as a case study.

This presentation will be held in 2036 Palmer Commons. There will also be a remote viewing option via Zoom.