Despite encouraging progress made by integrating multiplatform data for regulatory network reconstruction, identification of transcriptional regulatory networks remains challenging due to imperfection in current biotechnology and complexity of biological systems. It is important to develop new computational approaches for reliable regulatory network reconstruction, especially those of robustness against noise in gene expression data and 'structural error' (i.e., false connections) in binding data. We propose a new method, namely probabilistic network component analysis (pNCA), to estimate the posterior binding matrix given observed gene expression and binding data. The elements in the binding matrix, instead of taking deterministic binary values, are modeled as unknown Bernoulli random variables that represent the probability of regulation. A novel two-stage Gibbs sampling framework is employed to iteratively estimate both hidden transcription factor activities and the posterior distribution of binding matrix. Numerical simulation on synthetic data has demonstrated improved performance of the proposed method over several existing methods for regulatory network identification. Notably, the robustness of pNCA against 'structural error' in initial binding data is fortified with high tolerance of false negative connections in addition to that of false positive connections. The proposed method has been applied to breast cancer cell line data to reconstruct biologically meaningful regulatory networks, revealing condition-specific regulatory rewiring and important cooperative regulation associated with estrogen signaling and action in breast cancer cells.