Abstracting the dynamics of biological pathways using information theory: a case study of apoptosis pathway

Abstract

Supplementary data are available at Bioinformatics online.

Deep Abstractions of Chemical Reaction Networks

“…It is common enough to choose a projection over a subset of chemical species as abstraction function p -as in [26], possibly identified by information theoretic criteria to be those most influencing the reward of interest. Alternatively, we could follow [13] and use a projection over a certain number of sub-regions of the original state space in order to get a finite abstract state spaceS.…”

“…Related work. This idea was employed by Liu et al in [13] to approximate an ODE dynamic and was further refined by Palaniappan et al in [26] to deal with a stochastic dynamics. In this work, the authors select a subset of relevant variables and discretise them using information theoretic tools, and then build an approximate model based on a dynamic Bayesian network (DBN).…”

“…Contributions In this paper, we focus on the abstraction procedure, starting from [26]. Their approach is mostly driven by information theoretic considerations, which is extremely e↵ective in isolating the core components needed to perform an accurate approximation of the original process.…”

“…Their approach is mostly driven by information theoretic considerations, which is extremely e↵ective in isolating the core components needed to perform an accurate approximation of the original process. Moreover, [26] and [13] models are discrete in time and space: the domain of each tracked chemical species is partitioned in a certain number of subintervals and, at each time step, the system may change its state by jumping to another node of the discrete state space. The jumping probabilities, properly factored according to the DBN topology, have to be stored in memory.…”

Bayesian Abstraction of Markov Population Models

“…The final model is a Continuous Time Markov Chain (CTMC) where transition rates are learned using Gaussian Process Regression from some simulations of the full model. On the other hand of the spectrum, we find the work of Palaniappan et al [16], in which the authors start from a bunch of simulations of the original model, using information theoretic ideas extract a subset of relevant variables, discretize them and then learn a dynamic Bayesian network in discrete time. The abstract model was used for fast approximate simulation of the original model.…”