SParSE++: improved event-based stochastic parameter search

November 25, 2016

Abstract: 

Background
Despite the increasing availability of high performance computing capabilities, analysis and characterization of stochastic biochemical systems remain a computational challenge. To address this challenge, the Stochastic Parameter Search for Events (SParSE) was developed to automatically identify reaction rates that yield a probabilistic user-specified event. SParSE consists of three main components: the multi-level cross-entropy method, which identifies biasing parameters to push the system toward the event of interest, the related inverse biasing method, and an optional interpolation of identified parameters. While effective for many examples, SParSE depends on the existence of a sufficient amount of intrinsic stochasticity in the system of interest. In the absence of this stochasticity, SParSE can either converge slowly or not at all.

Results
We have developed SParSE++, a substantially improved algorithm for characterizing target events in terms of system parameters. SParSE++ makes use of a series of novel parameter leaping methods that accelerate the convergence rate to the target event, particularly in low stochasticity cases. In addition, the interpolation stage is modified to compute multiple interpolants and to choose the optimal one in a statistically rigorous manner. We demonstrate the performance of SParSE++ on four example systems: a birth-death process, a reversible isomerization model, SIRS disease dynamics, and a yeast polarization model. In all four cases, SParSE++ shows significantly improved computational efficiency over SParSE, with the largest improvements resulting from analyses with the strictest error tolerances.

Conclusions
As researchers continue to model realistic biochemical systems, the need for efficient methods to characterize target events will grow. The algorithmic advancements provided by SParSE++ fulfill this need, enabling characterization of computationally intensive biochemical events that are currently resistant to analysis.

Figure 1 illustrates ensemble results from running SParSE (a) and SParSE++ (b) a problem specification on which SParSE++ achieved the highest computational savings, with gain i =48.5% and gain t =46.6%. For the 30 initial reaction rates, SParSE computed a total of 166 estimates (332 iterations), whereas SParSE++ computed only 122 estimates (171 iterations).