Efficient framework for automated classification of subcellular patterns in budding yeast.

Fluorescent-tagging and digital imaging are widely used to determine the subcellular location of proteins. An extensive publicly available collection of images for most proteins expressed in the yeast S. cerevisae has provided both an important source of information on protein location but also a testbed for methods designed to automate the assignment of locations to unknown proteins. The first system for automated classification of subcellular patterns in these yeast images utilized a computationally expensive method for segmentation of images into individual cells and achieved an overall accuracy of 81%. The goal of the present study was to improve on both the computational efficiency and accuracy of this task. Numerical features derived from applying Gabor filters to small image patches were implemented so that patterns could be classified without segmentation into single cells. When tested on 20 classes of images visually classified as showing a single subcellular pattern, an overall accuracy of 87.8% was achieved, with 2330 images out of 2655 images in the UCSF dataset being correctly classified. On the 4 largest classes of these images, 95.3% accuracy was achieved. The improvement over the previous approach is not only in classification accuracy but also in computational efficiency, with the new approach taking about 1 h on a desktop computer to complete all steps required to perform a 6-fold cross validation on all images.