Granularity-aware fusion of biological networks for information extraction

Current biomedical research approaches capture multiple data sources, fusing and integrating them to extract accurate information. While the purpose of fusion in the biomedical domain is to create a more holistic picture of biological reality, latent sensitivities of the fusion process often result in information loss. It has been shown that biomedical data fusion is sensitive to granularity dimensions, scales of semantic relationships across specificity and mereology. Low granularity data casts a wide net, while high granularity data has focus. When data fusion occurs between low and high granularities, these benefits counteract each other. In this study, we reexamine the union function as a basis for biomedical data fusions, via comparison to a granularity-aware function. This granularity-aware approach uses domain knowledge (low granularity) as a filter for expression relationships (high granularity). We use pancreatic cancer expression data and domain knowledge networks to test the fusion approaches. We support previous findings that granularity-unaware fusion allows domain knowledge to eclipse condition-specific data. In addition, we find that the granularity-aware approach tends to outperform both the union and non-fusion networks, resulting in higher information extraction scores. Further, the granularity-aware approach increases the network information extraction effect size between disease and normal networks, allowing for a more distinctive delineation between the two conditions.