Integrated Machine Learning of Metabolic Networks Applied to Predictive Technology

The pharmaceutical industry invests huge resources in testing new compounds in drug development, and the ability to predict the toxicity of potential drugs is a vital component of this. Vast amounts of data are being generated by a range of different genomic approaches, but more bioinformatics tools are needed to help the user extract the relevant information.

Future Deliverables
Software packages for predictive toxicology in the pharmaceutical industry with the potential for transfer to non-pharma applications requiring complex predictive capability

Professors Steven Muggleton, Michael Sternberg and Jeremy Nicholson from Imperial College London, are bringing together their complementary expertise in computational and structural bioinformatics with metabolic research, to develop a computing tool -‘Metalog’- for predicting toxicity that could radically transform the whole drug development process. ‘Metalog’ will be based on metabolic pathway modeling, and will not only aid the user in predicting the toxic effects of compounds, but will also generate easily accessible explanations for how the predictions have been derived.

1. Development of a software tool to provide largely automated quantisation of metabolites in complex biological NMR spectra. Application to data sets obtained from a number of toxicology studies to provide input suitable for machine learning algorithms and generated liver toxin data sets.

2. Development of a logical model of inhibition in metabolic pathways using abduction and induction and application of machine learning for enzyme classification rules from biochemical reaction descriptions.

3. Development of a method (SVLIP) combing support vector machines (SVMs) and inductive logic programming (ILP) successfully predict the toxicity of small molecules (patent application is in progress).

4. Development of a novel representation for metabolic networks using Bayesian networks has been shown to give accurate predictions. Further studies reveal that the structures are robust and valuable to understanding the system under toxin effects (applied to liver toxin data sets such as Hydrazine, ANIT, CCl4).

At the core of ‘Metalog’ is machine learning. An innovative approach is being taken by combining different forms of machine learning, namely Bayes’ nets, Inductive Logic Programming and Support Vector Machines, each of which brings its own particular strengths. The team hopes that by developing an integrated approach, it will be able to represent the complex information which is associated with metabolic networks – be it information about the physical and structural properties of the metabolites, the topology of the network, or the uncertainty associated with metabolic reactions.

Bio/Pharma Applications

  • Predictive toxicology in drug discovery
  • Prediction of missing elements in metabolic networks
  • NMR data analysis
  • KEGG data analysis

A patent application is currently in preparation and it is envisaged that this project will generate a range of software packages for license

Model development will utilise descriptions of metabolic networks from existing databases via current and new collaborations with industry. These will provide complementary experimental data about compounds of immediate pharmaceutical interest. Models will also be developed from new experimental data being generated by Professor Nicholson’s metabonomics laboratory which will ultimately be a detailed and accurate dataset describing the effects of a particular case-study – the toxin, hydrazine – on changes in metabolite concentrations in vivo.

Industrial Interaction
The project is seeking to extend its collaborations especially in areas outside the pharmaceutical arena requiring predictive modelling solutions

Scientifically, this project will impact significantly on the crucial area of ‘systems biology’. Metabolic networks are relatively well understood (compared with gene networks for example) and comparisons between the models and the data are already demonstrating that a high level description of systems activity does not necessarily require knowledge about every component reaction. Also the comparison of metabolic network data derived from experimental analysis with an existing network database has revealed some unexpected interactions which warrant further analysis.

Significant Achievement
Analysis of NMR data on hydrazine toxicity into an appropriate form for enabling Bayes’ nets learning.

For further details on this project please see the project website or contact the team leader Professor Stephen Muggleton Department of Computing, email: Tel: 020 7594 8259

You may also like...

Leave a Reply

Your email address will not be published. Required fields are marked *