Feature reduction for improved recognition of subcellular location patterns in fluorescence microscope images

The central goal of proteomics is to clarify the mechanism by which each protein in a given cell type carries out its function. Automated protein subcellular location determination by fluorescence microscopy can play an important role in fulfilling this goal. The subcellular location of a protein is critical to understanding its function because each subcellular compartment has a unique biochemical environment. We have previously shown that neural network classifiers using sets of numerical features computed from fluorescence microscope images were able to recognize all major subcellular location patterns with reasonable accuracy. Current classifiers are limited by under-determined classification boundaries due to the limited number of available images compared to the number of features. In this paper, we compare various feature reduction methods that can address this problem. Specifically, principal component analysis, kernel principal component analysis, nonlinear principal component analysis, independent component analysis, classification trees, fractal dimensionality reduction, stepwise discriminant analysis, and genetic algorithms are used to select feature subsets that are evaluated using support vector machine classifiers. The best results were obtained using stepwise discriminant analysis and we found that as few as eight features can provide good classification accuracy for all major subcellular patterns in HeLa cells.