4. Conclusions
An analysis of data integration for quantitative transcriptional GRN modelling was presented, with a view toward investigating strategies to enhance network model quality and interpretation. A 27-gene network of transcriptional interactions was analysed. We used, as the gold standard, 16 interactions extracted from the DROID database, together with a time series gene expression dataset for the evaluation of simulation abilities. Integration was performed in two stages: exploring the interaction space (NSEx) and evaluating model behaviour (NSEv), through an evolutionary algorithm. The final objective was to provide a robust integration strategy in order to determine the best data type contributions to model performance at each stage.
The qualitative transcriptional gene interaction information was affected most by the integration of different data types in the NSEx stage of the algorithm, while integration at the NSEv stage affected both interaction identification and the ability to simulate continuous gene expression levels. This suggests that staged evaluation integration (NSEv) is mandatory if improved quantitative performance is desired. Importantly, NSEv was considerably more sensitive to input data type, reflecting the more stringent integration criterion mechanism. While exploration (NSEx) seemed to benefit from the consideration of all data, even those with less clear information on direct gene interactions, NSEv performed best when only binding site affinity data were integrated with the basic series. This suggests that even quite noisy data can be used to drive the search the scope of different GRN models, but that evaluation of the model set obtained requires better precision, i.e., more reliable data. The analysis presented here provided us with an optimal integration strategy, which, when applied to both time series datasets (SC and DC), led to further improvement in gene interaction information.
The general methodology for the data integration presented is applicable to any process driven by gene expression. However, it has been tested only on Drosophila melanogaster embryo development and associated datasets available for this system. When changing the system under analysis, e.g., to explore a different process or organism, both data types available and their quality may also change, so a similar staged analysis is crucial to determining the optimal integration strategy.