Subtractive clustering analysis: a novel data mining method for finding cell subpopulations

A novel data mining program called “subtractive clustering” picks out the most important differences between two or more flow cytometry listmode data files. While making no assumptions about the data, the program uses a variable weight and skew metric in the determination of bin size allowing for subtractive clustering of data without the need for bit-reduction or projection. In contrast, other subtraction methods, such as channel-by-channel subtraction, are dependent upon dimensionality and resolution, which can lead to an overestimation of positive cells because they do not account for the overall distribution of the test and control data sets. By taking into account human visual inspection of the data it is possible for the experimenter to choose an optimal subtraction by choosing an appropriate weight and skew metric, but without allowing direct modification of the results. By maximizing a bin size which can still differentiate clusters, it is possible to minimize computation while still removing data. The choice of control weight allows for different levels of bin destruction during the subtraction stage, the smaller the number the more conservative the subtraction, the larger, the more liberal. Three data sets illustrate full dimensional subtraction, single step biological data and multi-stage subtraction to show definitive test results. Subtractive clustering was able to conservatively remove control information leaving populations of interest. Subtractive clustering provides a powerful comparison of clusters and is a first step for finding non-obvious (hidden) differences and minimizing human prejudice during the analysis.