In this paper we survey work being conducted at Imperial College
on the use of machine learning to build Systems Biology models of the effects
of toxins on biochemical pathways. Several distinct, and complementary modelling
techniques are being explored. Firstly, work is being conducted on applying
Support-Vector ILP (SVILP) as an accurate means of screening high-toxicity
molecules. Secondly, Bayes' networks have been machine-learned to provide
causal maps of the effects of toxins on the network of metabolic reactions within
cells. The data were derived from a study on the effects of hydrazine toxicity in
rats. Although the resultant network can be partly explained in terms of existing
KEGG (Kyoto Encyclopedia of Genes and Genomes) pathway descriptions, several
of the strong dependencies in the Bayes' network involve metabolite pairs
with high separation in KEGG. Thirdly, in a complementary study KEGG pathways
are being used as background knowledge for explaining the same data using
a model constructed using Abductive ILP, a logic-based machine learning
technique. With a binary prediction model (up/down regulation) cross validation
results show that even with a restricted number of observed metabolites high
predictive accuracy (80-90%) is achieved on unseen metabolite concentrations.
Further increases in accuracy are achieved by allowing discovery of general rules
from additional literature data on hydrazine inhibition. Ongoing work is aimed
at formulating probabilistic logic models which combine the learned Bayes' network
and ILP models.
pubs.doc.ic.ac.uk: built & maintained by Ashok Argent-Katwala.