Advances in Bioinformatics

Review Article

A Survey of Flow Cytometry Data Analysis Methods

Table 1

Summary of survey (M: manual; Y: yes; E: embedded in gating; U: unsupervised; S: supervised; “—”: not supported, not implemented, not applicable; “”: same as above). Note that this table does not report Quality Assessment, Normalization, and Feature Extraction components.


Paper	Outlier removal	Automated gating				Labelling	Interpretation (classification/ comparison of samples)
		Method	Supervised/ Unsupervised	Multidimensional	Automated # of clusters

[45]	Logical and cleaning morphological operators applied to the corresponding image representation of FCM data	Logical operation on image representation of FCM data followed by thickening	U	—	—	Based on location and abundance of populations	—
		Majority operator applied to the image representation of FCM data followed by Soble edge detection	U	—	—		—
		Zero-degree B-Spline smoother applied to the 2-dimenisonal FCM data followed by break point detection	U	—	—		—
		Gath-Geva fuzzy clustering	U	—	—		—

[30]	Embedded in clustering (cluster membership weights can be used to exclude outliers)	Gaussian Mixture Models	U	Y		M	—
					Y
					(using BIC)


[46]	Embedded in clustering (cluster membership weights can be used to exclude outliers)	t-Mixture Models	U	Y		M	—
					Y
					(using BIC)


[47]	Embedded in clustering (excluding events that are far from Gaussian functions centers using a predefined cutoff value)	Mahalanobis distance from centroids of multivariate Gaussian functions used for classification task	S	—	—	E	—

[48]	—	Multilayer perceptron (MLP)	S	Y	—	E	—

[49]	—	Building templates for automated gating by using a cluster-finding algorithm (Beckton Dickinson’s (BD) snap-to gate algorithm)	U	—	—	E (initially set by operator)	—

[50]	—	DKLL (an extension of the -means algorithm to allow for non-spherical clusters)	U	Y	—	—	—
	—	Fuzzy -means based on adaptive distance	U	Y	—	—	—
	—	Fuzzy -means based on maximum likelihood	U	Y	—	—	—
	—	Fuzzy -means based on minimum total volume	U	Y	—	—	—
	—	Fuzzy -means based on sum of all normalized determinants	U	Y	—	—	—

[51]	—	M	—	—	—	M	Complete linkage hierarchical clustering

[52]	—	—	—	—	—	—	Comparing sample to a reference sample by probability binning algorithm

[53]	—	-means	U	Y	Histogram feature guided	—	—
					Partition index guided		—

[17]	—	Frequency difference gating approach (defines a gate(s) that contains statistically significant more events in the test sample than the control sample)¹	U	Y	—	—	—

[54]	—	MLP	S	Y	—	E	—
	—	Learning vector quantization (LVQ)	S	Y	—	E	—
	—	Radial basis function (RBF)	S	Y	—	E	—
	—	Asymetric RBF	S	Y	—	E	—
	—	Classification by modeling each class with Gaussian distributions	S	Y	—	E	—
	—	-nearest neighbour method	S	Y	—	E	—
	—	Kohonen’s self organizing map (SOM)	U	Y	—	M	—

[55]	—	Static gates applied to data	U	—	—	E (initially set by operator)	CLASSIF1 approach [56, 57]

[36]	—	Building templates for automated gating by using a cluster-finding algorithm (BD Snap-to gate algorithm)	U	—	—	E (initially set by operator)	—

[43]²	—	M	—	—	—	M	Functional linear discriminant analysis

[58]	—	Building templates for automated gating by using a cluster-finding algorithm (BD’s snap-to gate algorithm)	U	—	—	E (initially set by operator)	—

[59]	—	Gaussian Mixture Models	U	—	M	M	—

[60]	—	M	—	—	—	M	Average-linkage hierarchical clustering

[61]	—	M	—	—	—	M	Classification based on a semantic network of knowledge base through a hierarchical tree (if-then rule mechanism)

[62]	—	-means	U	Y	—	M	—
	—	Calculating modes of density function (calculated by Kernel density estimation ) followed by nearest neighbour heuristic	U	Y	—	M	—
	—	Gaussian mixture models using Markov chain Monte Carlo (MCMC)	U	Y	—	M	—

[63]	—	Building templates for automated gating by using a cluster-finding algorithm (BD’s snap-to gate algorithm)	U	—	—	E (initially set by operator)	—

[64]	—	Automated gating using BD Simulset software	—	—	—	M	Correlation tests using Spearman’s method

[65]	—	Image representation of randomly selected events from a group of flow data followed by smoothing, regional maxima detection and watershed algorithm to define the gates to apply to all the data	U	—	—	—	—

[66]	—	SOM	U	—	—	M	—
	—	Cluster analysis with Winlist (Verity Software House, USA))	U	—	—		—

[67]	—	Static gates applied to data and self adjusting gates (details not mentioned) for lymphocytes, monocytes, and granulocytes	U	—	—	E (initially set by operator)	CLASSIF1 approach [56, 57]

[68]	—	Fcom tool (an analysis tool in Winlist (Verity Software House, USA))	—	—	—	M	Average- linkage hierarchical clustering

[69]	—	Static gates applied to data and self adjusting gates for lymphocytes, monocytes, and granulocytes	U	—	—	E (initially set by operator)	CLASSIF1 approach [56, 57]

[70]	—	M	—	—	—	M	“Professor Fidelio” (a heuristic classification system that reasons on the basis of defined diagnostic patterns [71])

[72]	—	-means followed by Murphy’s cluster joining algorithm based on standard deviation of the data [73]	U	—	—	M	—
	—	-means followed by a cluster joining algorithm based on modified spread of the data and modified distance between two clusters [72]	U	—	—	M	—
	—	Preclustering a subset of the data by -means and assigning unclustered events to the closest cluster center followed by a cluster joining algorithm based on modified spread of the data and modified distance between two cluster [72]	U	—	—	M	—

[73]	E (excluding the events that were more than a set number of standard deviations away from the centroids of the clusters)	-means followed by Murphy’s cluster joining algorithm based on standard deviation of the data	U	Y	—	M	—

[74]	—	MLP	S	Y	—	E	—

[75]	—	RBF	S	Y	—	E	—

[76]	—	MLP	S	Y	—	E	—
	—	SOM	U	Y	—	M	—
	E (excluding the events that were more than a set number of standard deviations away from the centroids of the cluster)	-means	U	Y	—	M	—

[77]	—	No gating—mean fluorescent intensities of antibodies were used for next stage of analysis	—	—	—	—	MLP

[78]	—	RBF	S	Y	—	E	—

[40]	—	—	—	—	—	—	Histogram of one parameter of FCM data followed by MLP

[79]	—	Classification and regression trees (CARTs)	S	Y	—	E	—

[80]	—	Support vector machine (SVM)	S	Y	—	E	—
	—	RBF	S	Y	—	E	—

[81]	—	RBF using radially symmetric basis function (based on Euclidean distance)	S	Y	—	E	—
	—	RBF using more general arbitrarily oriented ellipsoidal basis functions (based on Mahalanobis distance)	S	Y	—	E	—

[82]	—	Gaussian mixture model clustering	U	Y	—	—	—

[83]	Embedded in clustering (excluding events that are far from Gaussian functions centers using a predefined cutoff value)	Mahalanobis distance from the centroids of multivariate Gaussian functions used for classification task	S	—	—	E	—

[84]	—	M	—	—	—	—	Classification based on a shrunken centroids approach [85]
							Hierarchical clustering

[41]	—	M	—	—	—	—	Kernel density estimation followed by calculating differences between patients by Kulback-Leibler divergence to form a similarity matrix and then dimensionality reduction by multidimensional scaling for 2-dimensional visualization

[86]	—	RBF	S	Y	—	E	—

[87]	—	M	—	—	—	—	Hierarchical clustering
							Principal component analysis (PCA) for dimensionality reduction and visualization to see if classes are separable by looking at the first few principle components

¹Closely related to probability binning algorithm introduced in [52].
²This study utilizes quality assessment strategy introduced in [42] that is based on comparison of density, ECDF (empirical cumulative distribution function), box plots, and two types of bivariate plots of similar samples.