 |
Center for Nonlinear Studies |
Individual cells within an isogenic population may exhibit widely varying responses to the same stimuli and identifying the biochemical mechanisms that account for heterogeneity is one of the major challenges in systems biology. I will talk about several computational approaches we have been developing to determine the mechanistic basis for individuality at the single-cell level using Bayesian parameter estimation (BPE) and related Monte Carlo approaches for sensitivity analysis. BPE is widely-used to learn models from data, but standard algorithms, such as Markov Chain Monte Carlo (MCMC), suffer from slow convergence when applied to complex models, particularly with sparse data as is the usual case for biological systems. We have used an accelerated sampling method called parallel tempering, which uses multiple Markov Chains run in parallel at different temperatures, to improve the performance of BPE for complex models. In addition, we have applied a sparsity-promoting penalty called Lasso, which is widely-used in machine learning, to reduce model complexity and to identify in a systematic fashion the mechanisms within a model that are required to reproduce key features of the data. Finally, we have applied a variant of Principal Component Analysis that we call “Last Component Analysis” (LCA) to identify the model variables that contribute maximally to a given property of interest. We show with several examples that LCA can identify combinations of parameters whose variation correlates strongly to a given property of interest, thus establishing potential mechanisms that can account for observed heterogeneity. **This seminar is part of a series on Artificial Intelligence for Computational Science.
Host: Aric Hagberg