Reading and Assembling Contextual and Holistic Big Mechanisms
Developers: Angus Forbes, M. Surdeanu, K. Barnard, C. Morrison, R. Gutenkunst, G. Yao
Funding: DARPA Big Mechanism, Grant #BAA-14-14
The REACH system extracts (i.e., reads and assembles) causal mechanisms from research abstracts and papers. The fundamental research proposed has three guiding principles. The mechanisms extracted must: (a) be runnable, i.e., they must predict the data; (b) model current uncertainty of causal relations being active given context information; and (c) be holistic, that is, they explain multiple layers of abstraction, ranging from cellular to organism phenotypes. Thus, running our model produces quantifiable probabilities for the modeled variables (e.g., presence of pancreatic cancer) given customizable input values (e.g., overexpression of the VEGF protein). We assume that the model is deterministic given a specified context, but, in practice, variance in results exists due to difference in context - either reported or completely unknown - in the acquired mechanisms. The REACH system will provide a platform for communicating explanation with novel visualization tools for the extracted mechanisms. We will evaluate our approach in the domain of cancer biology, by extracting signaling pathways, but our work is largely domain independent.
Date: July 1, 2014 - June 30, 2018