Proceedings of the XLVI Italian Society of Agricultural Genetics - SIGA Annual Congress

Giardini Naxos, Italy - 18/21 September, 2002

ISBN 88-900622-3-1

Poster Abstract - 4.29

BACKGROUND FLOURESCENCE IN cDNA MICROARRAY EXPERIMENTS

STEFANINI F.M., CAMUSSI A.

Dipartimento di Statistica “G.Parenti”, Viale Morgagni 59, and Facoltà di Agraria, Firenze

Dipartimento di Biotecnologie Agrarie, P.le delle Cascine 24, Firenze

microarrays, spatial models, background fluorescence, normalization

Typical cDNA microarray experiments are affected by several sources of noise. Background fluorescence is measured to estimate the amount of fluorescence due to coating molecules and, eventually, to dying molecules leaked from spots.

The normalization of spot fluorescence is often performed by subtracting the background fluorescence , XB, from the spot raw intensity, XR, and this step is performed before applying any other transformation to normalize array data (Quackenbush, 2001). Due to the heterogeneous nature of the background fluorescence, the correction XR - XB may result in a sub-optimal normalization procedure, for example because of dust that causes the overestimate of background noise components. Negative corrected values of fluorescence intensity are a straightforward consequence of over-correction.

A first extension of the background correction has been proposed by Stefanini (2002) by introducing two coefficients beta in a hierarchical Bayesian model, thus the correction (after simplification) resembles to X R - b _d XB. The posterior distribution of the dye-specific parameters are far below one thus evidences were found that the normalization may be improved in this respect.

Nevertheless, background data contain even more information then used in Stefanini(2002). For example, the correlation of fluorescence intensities among closely related array locations might be exploited to improve the quality of estimates. In several experimental protocols it has been observed that laser confocal microscopes may behave differentially according to the distance of a spot from the location of perfect focus. The confirmation of these findings would imply that a systematic source of noise might be removed from the background (and from spots) thus improving the signal-to-noise ratio, i.e. the quality of the analysis.

In this contribution, statistical models are used to smooth raw background values in order to improve the normalization of raw data and to test for the presence of a trend surface induced by sources of bias like the confocal microscope. Let D C R² be the two-dimensional array area in which n background values Z(x i ),i=1,...,n are observed from the process Z(x ),x E D. By assuming the presence of a trend surface, we have Z(x )=h(x )+ e (x ), where h(x )=constant if the trend is absent, and where e (x ) is a zero mean spatial stochastic process with pre-assigned covariance.

The proposed model is applied to an actual dataset from a dye-swap cDNA experiment. Results show that a suitable statistical model may improve the quality of the normalization, in particular by avoiding over-correction of spot raw fluorescence intensity.

Quackenbush, J.,2001, Computational analysis of microarray data, Nature Reviews Genetics, 2:418-427.

Stefanini F.M., 2002, A model-based normalization for cDNA microarray experiments, Società Italiana di Statistica, Congress Acta, Milano, June 2002.