Trend correlations (i.e., two genes are correlated in their varying trends that rise and descend with time) between genes are very important but usually neglected in reconstruction of gene regulatory networks (GRN). To mine trend correlations to enhance the reconstruction performance of GRN, we propose geometric-pattern dynamic Bayesian networks (Gp-DBN). In Gp-DBN the time series of each gene is transformed to a geometric pattern, by which potential regulators and time lags are estimated, and regulatory relations between genes are discovered by reasoning correlations between these geometric patterns. Gp-DBN realizes the mining of regulatory relations with trend correlations so that it can improve the performance of GRN reconstruction. Experimental results on Yeast and E. coli data sets show that Gp-DBN improves greatly the performance of GRN reconstruction in the cases with／without prior knowledge, compared with traditional dynamic Bayesian networks.
