Background: Factor analysis ( FA) has been widely applied in microarray studies as a data-reduction-tool without any a-priori assumption regarding associations between observed data and latent structure ( Exploratory Factor Analysis).
A disadvantage is that the representation of data in a reduced set of dimensions can be difficult to interpret, as biological contrasts do not necessarily coincide with single dimensions. However, FA can also be applied as an instrument to confirm what is expected on the basis of pre-established hypotheses ( Confirmatory Factor Analysis, CFA). We show that with a hypothesis incorporated in a balanced (orthogonal) design, including 'SelfSelf' hybridizations, dye swaps and independent replications, FA can be used to identify the latent factors underlying the correlation structure among the observed two-color microarray data. An orthogonal design will reflect the principal components associated with each experimental factor. We applied CFA to a microarray study performed to investigate cisplatin resistance in four ovarian cancer cell lines, which only differ in their degree of cisplatin resistance.
Results: Two latent factors, coinciding with principal components, representing the differences in cisplatin resistance between the four ovarian cancer cell lines were easily identified. From these two factors 315 genes associated with cisplatin resistance were selected, 199 genes from the first factor ( False Discovery Rate (FDR): 19%) and 152 ( FDR: 24%) from the second factor, while both gene sets shared 36. The differential expression of 16 genes was validated with reverse transcription-polymerase chain reaction.
Conclusion: Our results show that FA is an efficient method to analyze two-color microarray data provided that there is a pre-defined hypothesis reflected in an orthogonal design.