Visualizing biomarkers and their association with clinical outcomes: A machine learning approach

Clinical investigators often seek to identify biomarkers that are associated with clinical outcomes. The challenge is that biomarkers tend to be numerous, often inter-correlated, and with varying degrees of association with the clinical outcome of interest. We developed a straightforward, effective visualization approach that allows investigators to see the inter-correlations of biomarkers and their association with clinical outcomes rendered in a 2-dimensional (2-D) plot.

We extracted pairwise differences between outcome groups. Next, a t-SNE machine learning method was applied to reduce dimensionality of the data into 2-dimensional data. This data was then rendered as 2-D plots allowing visualization of either the original biomarker values or their ranking in the sample. We apply this method to a dataset of 235 subjects with data of 114 metabolomic biomarkers and cardiovascular outcomes.

The 2-D plots demonstrated that the metabolite biomarkers with high degrees of association with the outcome tend to be positioned away from non-significant markers. Markers that are correlated with each other are positioned in clusters. An online demo version application of this visualization tool is available at http://52.9.140.88:3838/MarkersVu/.

This is an effective approach to visualize biomarkers and their correlations, allowing rapid visual identification of biomarkers and marker clusters with potentially higher degree of association with the outcome.