Misty Mountain clustering: application to fast unsupervised flow cytometry gating

BMC Bioinformatics

Table 3 Summary of comparing Misty Mountain with state of the art flow cytometry specific clustering methods

Data set				Manually gated 2D barcoding^&		Simulated 5D Gaussians	Simulated 2D non-convex		3D rituximab	4D GvHD	Manually gated 4D OP9
Misty Mountain	accuracy	sens	(%)	100		100	100		-	-	100
		spec	(%)	100		100	100		-	-	100
	CPU	(sec)		10		196	6		0.3	0.8	3.6
FLAME	accuracy	sens	(%)	20^a	60^b	-	0^d*	100^d	-	-	-
		spec	(%)	33^a	50^b	-	0^d*	100^d	-	-	-
		CPU	(sec)	5.10⁴		>3.10⁵	1.10⁴		10	360	1.4 · 10⁴
flowClust	accuracy	sens	(%)	45^a*	60^b*	100^c	0^c*	100^d	-	-	60^d*	60*
		spec	(%)	60^a*	55^b*	100^c	0^c*	100^d	-	-	75^d*	38*
	CPU	(sec)		5.10⁴		4.10⁴	7200		43	480	3660
flowMerge	accuracy	sens	(%)	25		100	0		-	-	80
		spec	(%)	45		100	0		-	-	57
	CPU	(sec)		1.3 · 10⁵		1.27 · 10⁵	7200		124	1020	8400
flowJo	accuracy	sens	(%)	45		-	-		-	-	-
		spec	(%)	47		-	-		-	-	-
	CPU	(sec)		1-10		-	-		1-10	1-10	-

^a optimal cluster number: 12
^b optimal cluster number: 24
^a*optimal cluster number: 15
^b*optimal cluster number: 22
^c optimal cluster number: 5
^c* optimal cluster number: 2
^d optimal cluster number: 1
^d* optimal cluster number: 4
* optimal cluster number: 8
^&to save CPU time a data set, reduced by 80%, has been analyzed by FLAME, flowClust and flowJo
sens (sensitivity) = (# of correctly assigned clusters)/(# of clusters in gold standard)
spec(specificity) = (# of correctly assigned clusters)/(total # of assigned clusters)
Gold standards were independent expert manual clustering for experimental data and specified clusters for simulated data.

ISSN: 1471-2105