Background : Medicine systems disease maps (DM) summarize knowledge about molecular pathophysiology into process description networks (PDN). PDN provide a network representation wherein molecules states are described and linked to processes they are involved in, and processes are linked together by the molecules they share. Direct interaction networks (DIN) are networks based on direct interactions between unprocessed molecules. The choice of the network type is a key element in systems biology analysis. PDN may be too complex for some analysis, but also the only knowledge resource available. DM provide a fine representation of molecular reactions, including information for metabolic reactions, transduction signaling and gene expression regulation. Thus, we hypothesize that DIN could be deduced from information provided by DM. More specifically, ontologies, which are able to manage knowledge and relationships using logical rules, are a suitable framework to deal with this large-scale inference. We present an ontology-driven methodology that results in the deduction and the addition of DIN relationships to PDN ones. Methods : We designed the Molecular Network Ontology (MNO), which contains 42 classes that represent a) molecular reactions (i.e. binding, conversion or transcription) and b) molecular participants (e.g. gene, native gene product or converted gene product). Process classes were formally defined according to participant classes using 4 PDN relationships: "has input", "has output", "positively mediated by" or "negatively mediated by". Then, based on these relationships and process classes, logical rules were designed to infer DIN relationships (e.g. "symmetrically interact with" or "positively directed to"). Use Case : The macrophage signal transduction map (MSTM) is a curated DM that contains 724 molecular reactions involving 1,353 participants. MSTM network was integrated into MNO: reactions and participants described in MSTN became individual instances of MNO classes. Edges from MSTN became PDN relationships between instances. Meta-information (labels, identifiers, crosslinks…) were kept as individual annotations. Following automatic reasoning, the patterns of MNO-inferred DIN relationships and of those provided by STRING queries were compared to validate the consistency of deductions. Results : MNO could fully integrate MSTM information as individual instances of its classes. Then logical rules allowed us to enrich the initial PDN relationships with 36,442 consistent DIN relationships. Conclusion : Ontologies are an adequate framework to manipulate networks. Subsequent logical reasoning benefits from the genericity of a systems approach, leading to the inference of consistent DIN new relationships from PDN. Finally, such a methodology opens perspective to expand the choice of appropriate networks for systems biology analysis.