Model guided trait-specific co-expression network estimation as a new perspective for identifying molecular interactions and pathways

A wide variety of parametric approaches and co-expression networks have been developed for finding gene-by-gene interactions underlying complex traits from expression data. However, a little is known about the practical correspondence and synergistic potential of these different schemes. We provide a framework for parallel consideration of parametric interaction models with quantitative traits and co-expression networks based on a previously uncharacterized link between them. Resulting trait-specific co-expression network estimation method 1) serves to enhance the interpretation of biological networks in a more parametric sense and 2) exploits the underlying parametric model itself in the estimation process. It is tailored for simultaneous identification and classification of molecular interactions and pathways regulating complex traits by accounting for common characteristics of genetic architectures due to which the mainstream methods often lack efficiency. A remarkable advance over the state-of-art methods is illustrated theoretically and through comprehensive simulated scenarios. In particular, prognostically important novel findings in acute myeloid leukemia analysis demonstrate the methods immediate practical relevance. Author summaryHere we built up a mathematically justified bridge between parametric approaches and co-expression networks that have become prevalent for identifying molecular interactions underlying complex traits. We first shared our concern that methodological improvements around these schemes adjusting only their power and scalability are bounded by more fundamental scheme-specific limitations. Subsequently, our theoretical results were exploited to overcome these limitations to find gene-by-gene interactions neither of which can capture alone. We also aimed to illustrate theoretically and empirically how this framework enables the interpretation of co-expression networks in a more parametric sense to achieve systematic insights into complex biological processes more reliably. The main procedure was fit for various types of biological applications and high-dimensional data to cover the area of systems biology as broadly as possible. In particular, we chose to illustrate the methods applicability for gene-profile based risk-stratification in cancer research using public acute myeloid leukemia datasets.