During the last years gene interaction networks are increasingly being used for the assessment and interpretation of biological measurements. Knowledge of the interaction partners of an unknown protein allows scientists to understand the complex relationships between genetic products, helps to reveal unknown biological functions and pathways, and get a more detailed picture of an organism's complexity. Being able to measure all protein interactions under all relevant conditions is virtually impossible. Hence, computational methods integrating different datasets for predicting gene interactions are needed. However, when integrating different sources one has to account for the fact that some parts of the information may be redundant, which may lead to an overestimation of the true likelihood of an interaction. Our method integrates information derived from three different databases (Bioverse, HiMAP and STRING) for predicting human gene interactions. A Bayesian approach was implemented in order to integrate the different data sources on a common quantitative scale. An important assumption of the Bayesian integration is independence of the input data (features). Our study shows that the conditional dependency cannot be ignored when combining gene interaction databases that rely on partially overlapping input data. In addition, we show how the correlation structure between the databases can be detected and we propose a linear model to correct for this bias. Benchmarking the results against two independent reference data sets shows that the integrated model outperforms the individual datasets. Our method provides an intuitive strategy for weighting the different features while accounting for their conditional dependencies.