Many diseases with genetic background are not just caused by single genes, but by a complex interplay between several genes and environmental factors. Thus, the investigation of their interactions gains more and more importance. However, statistical modeling of gene-gene interactions is a challenge. For example, regression models do not fully capture biological interaction. We may look for statistical approaches that offer more flexibility. Artificial neural networks do not depend on pre-specified model structures and may thus be more appropriate to identify biological independence or interaction. Our approach is based on the idea that neural networks with reduced topology, i.e., a topology where the independent loci are not connected, should be able to reflect a biological independence model. Thus, we compare the model fits of two neural networks one with a reduced and one with a fully connected topology which should allow to decide on the presence of biological interactions. We perform a simulation study to investigate whether our approach leads to satisfactory results assuming different biological models of independence and interactions. It can be concluded that obviously similar problems as with standard regression models occur such that our approach in its present form is not able to identify biological independence.

artificial neural networks, biological independence, gene-gene interaction, two-locus disease model.