Error-tracking clustering gives quantitative statistics to DNA segmentation analysis

Inferences acquired by applying clustering analysis of microarrays cannot be reliably assessed before data-originated errors are quantified, an exacting task that is often not performed. Here, we present a novel and fast clustering technique, pair-wise Gaussian merging (PGM), suited for this purpose. Designed for systems with normally distributed error, PGM treats each observation as a Gaussian distribution function, with error as width, uses a simple but exact mathematical relation to track error at every step of clustering, and gives results from which quantitative statistics are easily extracted. PGM is built on a framework of agglomerative hierarchical clustering, uses t-value as distance and requires no linkage criteria. We demonstrate the merits of PGM by applying it to a segmentation algorithm for DNA copy number analysis (SAD) which, by comparing it’s performance to existing algorithms, we verify that it: provides quantitative statistics for predictions; is simpler in formulation; is less thirsty for memory; offers higher accuracy; and for today’s typical array size, is faster by orders of magnitude than its nearest competitor. With only two user-adjusted and easily comprehended parameters, SAD is highly user friendly. SAD’s running time scales linearly with data size and is therefore ideally suited to the challenge of ever-growing array resolution. On a typical modern notebook, SAD completes high-quality copy number analysis for a 250 thousand-marker array in ∼1 second and a 1.8 million-marker array in ∼8 seconds.