IC Crest   DTI Harnessing Genomics Beacon Project:

MetaLog - Integrated Machine Learning of Metabolic Networks applied to Predictive Toxicology

DOC Home
CBL Home >  Metalog Home
[On-Line Resources] [Poster]


The goal of this research programme is to develop an industrial-strength tool, MetaLog, for modelling disease pathways and predicting toxicity using machine learned models of proteins and metabolic activity.  MetaLog will be based on a combination of machine learning and uncertainty reasoning. MetaLog will be developed around problems involving pathway modelling and predicting the toxicity of a variety of compounds of interest to the industrial partners. Predictive Toxicology is a vital area for the pharmaceutical industry (related to Safety Assessment) and can be viewed as an exemplar for post-genomic biomedical research. In the pharmaceutical sector predictive toxicology is important for cutting R&D costs and reducing the attrition rate of compounds in development. Repeatedly in bioinformatic modelling the best algorithm often is an ad hoc amalgam of different machine learning algorithms. This results in the end-user being unable to understand the reasoning involved in predictions. We therefore identify a need both specifically for modelling of metabolic activity and generally for bioinformatics based on developing a formal approach to combine different machine learning approaches and to make the reasoning accessible to end-users.

The project objectives are:
  1. to develop a formal framework to combine different machine learning approaches, namely Bayes' nets and Inductive Logic Programming (ILP) and to make this reasoning methodology accessible to pharmaceutical end users;
  2. to use the above developments together with metabonomic and proteomic data to develop an industrial-strength tool, MetaLog, which automatically hypothesises missing elements of metabolic networks and provides explanations and predictions concerning the toxicity of a variety of compounds;

Principal Investigators:

In partnership with: