Reconstruction of gene interaction networks from experimental data provides a deep understanding of the underlying biological mechanisms. The noisy nature of the data and the large size of the network make this a very challenging task. Complex approaches handle the stochastic nature of the data but can only do this for small networks; simpler, linear models generate large networks but with less reliability.

We propose a divide-and-conquer approach using probabilistic graph representations and external knowledge. We cluster the experimental data and learn an interaction network for each cluster, which are merged using the interaction network for the representative genes selected for each cluster.

We generated an interaction atlas for 337 human pathways yielding a network of 11,454 genes with 17,777 edges. Simulated gene expression data from this atlas formed the basis for reconstruction. Based on the area under the curve of the precision-recall curve, the proposed approach outperformed the baseline (random classifier) by ∼15-fold and conventional methods by ∼5–17-fold. The performance of the proposed workflow is significantly linked to the accuracy of the clustering step that tries to identify the modularity of the underlying biological mechanisms.

We provide an interaction atlas generation workflow optimizing the algorithm/parameter selection. The proposed approach integrates external knowledge in the reconstruction of the interactome using probabilistic graphs. Network characterization and understanding long-range effects in interaction atlases provide means for comparative analysis with implications in biomarker discovery and therapeutic approaches. The proposed workflow is freely available at http://otulab.unl.edu/atlas.