Inferring Networks from Biological Intuition

摘要

What do data from large-scale, parallel physiology (LSP) experiments such as microarrays and pro-teomics mean? The original hope for these experiments was that they could be used to unravel the mechanisms that produce phenotypes of interest, such as diseases and responses to drugs. Better ways to determine the mechanisms of multifac-torial and variably penetrant diseases are badly needed, especially for phenotypes such as the dementias that are both devastating and very difficult to treat. Though there Has been some progress in inferring small networks (1-5 molecular species) using different statistical methods, the use of LSP data to determine mechanisms remains an elusive goal. Instead, LSP experiments are used today to improve the pathological classification of specimens by determining which molecular observations best correlate with disease phenotypes and to identify genes and gene products that might be related to a phenotype in some way. There are very sound theoretical and practical reasons for this change, ranging from the combinatorial explosion of possible mechanisms and statistical issues to the impossibility of certain experiments and high biological variation.rnNonetheless, biologists frequently infer mechanisms from very fragmentary and imperfect data. How they do tnis provides a clue to another approach to causal inference. After candidate genes and gene products are identified from a microar-ray experiment by one or more ad hoc methods, the biologist tests each candidate by asking somernpertinent questions. In this paper we explore the computational meanings of these questions and consider how their implementations could produce mechanistic inferences.