An analysis of extensible modelling for functional genomics data

In this paper we review the data formats that exist in functional genomics, some of which have become de facto or de jure standards, with a particular focus on how each domain has been modelled, and how each format allows extensions. We describe the tasks that are frequently performed over data formats and analyse how well each task is supported by a particular modelling structure.From our analysis, we make recommendations as to the types of modelling structure that are most suitable for particular types of experimental annotation. There are several standards currently under development that we believe could benefit from systematically following a set of guidelines.The advent of large scale approaches investigating biological systems has generated a requirement for standard data formats that has been recognised by the bioinformatics community for several years. It is a major challenge to create standards that are stable and "future proof" for considerable lengths of time. In this document, we review the models associated with standard data formats for microarrays, proteomics and metabolomics (collectively known as functional genomics). The experimental techniques in these areas are evolving rapidly, different laboratories use different instruments and software, and a single experiment can produce a wide range of heterogeneous data types. This causes problems because data produced in one laboratory often cannot be interpreted by other groups or compared with other data sets produced in a different setting. Proposals have been made for data standards for microarrays (MAGE-ML [1]), protein-protein interactions (the Molecular Interaction format [2]), mass spectrometry (most recently mzData [3] and mzXML [4]), and protein separation proteomics (PEDRo [5]). There have also been proposed extensions to MAGE-ML to accommodate other types of experiment (FGE-OM [6] and SysBio-OM [7]). Data standards for metabolomics are at an early stage, but there are three models that could