Detecting a difference - assessing generalisability when modelling metabolome fingerprint data in longer term studies of genetically modified plants

Type Article
Original languageEnglish
Pages (from-to)335-347
Number of pages13
JournalMetabolomics
Volume3
Issue number3
Early online date19 Jun 2007
DOI
Publication statusPublished - Sep 2007
Links
Permanent link
View graph of relations
Citation formats

Abstract

There is current debate on whether genetically-manipulated plants might contain unexpected, potentially undesirable, changes in overall metabolite composition relative to that of the progenitor genotype. However, appropriate analytical technology and acceptable metrics of compositional similarity require development, particularly to allow data integration from different laboratories and different harvests. For an initial comprehensive overview of compositional similarity, we explored the use of a rapid and relatively non-selective fingerprinting technique based on flow injection electrospray ionisation mass spectrometry (FIE-MS). Six conventionally-bred potato cultivars and six experimental bioengineered potato genotypes were produced in four field blocks during two growing seasons and analysed on two different analytical instruments (LCT, Micromass in 2001 and LTQ, Thermo Finnigan in 2003). Field effects and overall process variability was found to be negligible when compared to inherited genotype variance. The data derived separately for experiments using tubers from individual harvest years were compared to assess the generalisability of models for the comparison of GM and non-GM potato tubers under investigation. This procedure proved appropriate for not only rapid assessment of similarities between plant genotypes but also to predict the identity of metabolite signals that could explain differences between genotype classes irrespective of the instrument used for analysis. Importantly, despite differences in ionisation and data acquisition properties of the two instruments the generalisation of models could be confirmed after correlation analysis of explanatory variables correctly identified the molecular origin of differences between genotypes. We conclude that FIE-MS metabolomics fingerprinting technology coupled to machine learning data analysis has great potential as a robust tool for first-pass metabolic phenotyping and, therefore, initial assessments of compositional similarities prior to use of more targeted hyphenated gas or liquid chromatography-mass spectrometry techniques.