An Improved Distance Metric For Clustering Gene Expression Time-Series Data
Philip Heller, Bharath Baiju · Zenodo (CERN European Organization for Nuclear Research) · 2018
Gene expression in cells can fluctuate over time in response to internal or external stimuli. Time series expression studies can provide much information about how organisms function and respond to the environment. Due to economic constraints, such studies usually contain few time points and many genes; therefore, traditional time-series analysis techniques are not applicable. Genes with similar expressions patterns generally share common function. Unsupervised clustering of genes on the basis of expression pattern can provide valuable insight functional relationships among genes. This insight is particularly valuable in the case of environmental bacteria, where the function of most genes of most species is unknown. Clustering requires a distance metric to quantify the pairwise differences among objects being clustered. In the case of gene expression, the most common metric is Pearson’s Correlation Coefficient (PCC). Despite its popularity, PCC has a number of drawbacks: it does not match up with intuitive notions of distance, and it is insensitive to timepoint ordering. Consequently, clusters computed on the basis of PCC can contain dissimilar members and can lead to erroneous conclusions. We propose a new metric, called ABLIM (“Area Between Linear Interpolations of Measurements”), which overcomes the shortcomings of PCC. ABLIM is visually intuitive, obeys triangle inequality, and is sensitive to timepoint ordering. Comparison of ABLIM and PCC clustering of gene expression data for the marine bacterium Crocosphaera watsonii demonstrates that ABLIM-based clusters more reliably reflect biological reality.