Title: Sensitivity biplots
Authors: Julia Fukuyama - Indiana University (United States) [presenting]
Abstract: In high-dimensional biological datasets, distances customized to the biological system being studied are often used in conjunction with multi-dimensional scaling in order to obtain a lower-dimensional representation of the data. This low-dimensional representation can be used either for exploratory analysis or to obtain a smaller set of variables for later statistical analysis. Multi-dimensional scaling has an advantage over linear methods such as PCA in that it can be used to embed any set of distances between samples, but it has the disadvantage of being difficult to interpret in terms of the original variables. To overcome this limitation, we present a generalization of PCA biplots to multi-dimensional scaling. These plots describe the relationship between the multi-dimensional scaling embedding space and the original variables, allowing a visualization of variable importance in different parts of the space. We illustrate the method on a human microbiome dataset, showing how it gives insight into both the distance used to construct the embedding and the relevant biology.