Clustering-based methodology with minimal user supervision for displaying cell-phenotype signatures in image-based screening
Abstract
Most quantitative cell image-based screening analyses are dependent on thorough user supervision based on assay-specific knowledge. To minimize human bias in analysis, we introduce an automated methodology of displaying screen phenotypes using clustering that provides intuitive visuals to guide user supervision when required. Our premise is to automatically present to users an overview of screen phenotype-contents to assist in planning assay and analysis of a new screen. Our methodology starts from numerical features of cell-images, removes outliers through the density-based clustering OPTICS, identifies significant phenotypes by the Hierarchical Agglomerative Clustering and Dynamic Tree Cut techniques, and displays representative cell images as phenotype-signatures. User supervision needed to detect outliers and adjust the desired heterogeneity-level of identified phenotypes is facilitated respectively by an intuitive density plot and a systematic phenotype display. The methodology was tested on various phenotypes of the Golgi apparatus, an intracellular structure essential for cell physiology and protein secretion. The Golgi apparatus was targeted by various drug treatments. This test demonstrates the methodology's potentials by providing a comprehensive categorization of Golgi phenotypes.
Authors: William-Chandra Tjhi, Kee Khoon Lee, Terence Hung, Yew-Soon Ong, Ivor Wai-Hung Tsang, Victor Racine, Frédéric Bard
Published in: IEEE International Conference on Bioinformatics and Biomedicine Workshops (BIBMW) (2010)