Automatic identification of subcellular phenotypes on human cell arrays

Abstract

Light microscopic analysis of cell morphology provides a high-content readout of cell function and protein localization. Cell arrays and microwell transfection assays on cultured cells have made cell phenotype analysis accessible to high-throughput experiments. Both the localization of each protein in the proteome and the effect of RNAi knock-down of individual genes on cell morphology can be assayed by manual inspection of microscopic images. However, the use of morphological readouts for functional genomics requires fast and automatic identification of complex cellular phenotypes. Here, we present a fully automated platform for high-throughput cell phenotype screening combining human live cell arrays, screening microscopy, and machine-learning-based classification methods. Efficiency of this platform is demonstrated by classification of eleven subcellular patterns marked by GFP-tagged proteins. Our classification method can be adapted to virtually any microscopic assay based on cell morphology, opening a wide range of applications including large-scale RNAi screening in human cells.

Figure 1

Workflow of phenotype classification system. The mix of targeted DNA (e.g., subcellular clones–GFP fusion), gelatin, and transfection reagents are prepared in microwells and spotted in arrays into chamber slides by a DNA-spotting robot. Cultured cells are added to the chamber slides and transfected with the DNA on the arrayed spots. After 10 h incubation at 37°C the fluorescence signal of the expressed GFP signal can be visualized. On the automatic scanning platform a motorized stage and motorized z-stepper perform the scanning of the live cells. The control of the illumination and the automatic calculation of the integration time allow an automatic acquisition of cell images. Cells with GFP-signals are captured and image features are extracted. These features serve as input for training of the automated classification system.

Figure 2

Automatically captured images of live MCF7 cells representing localization classes. (A) Cells were counterstained by Hoechst (blue) and different GFP-tagged cDNAs were expressed in cells transfected on cell arrays (green). Image acquisition is designed in such a way that each image contains only one cell. From top left to bottom right: GFP-tagged NES→cytoplasm (96 images), EGFP→cyto-nuclear (170 images), GFP-tagged cDNA #488→mitochondria (108 images), YFP-tagged LB1→nuclear lamina (108 images), GFP-tagged ErbB1→plasma membrane (70 images), YFP-tagged SRb→endoplasmatic reticulum (84 images), GFP-tagged cDNA #351→nucleoli (116 images), YFP-tagged cDNA #447→peroxisomes (119 images), YFP-tagged cDNA #22f21→microtubules (78 images), YFP-tagged H2B→nuclear (111 images), YFP-tagged GalT→Golgi (93 images) imaged from two different spatial directions. For cDNA reference see Supplemental Table 1 (see also http://www.dkfz.de/LIFEdb/ and http://harvester.embl.de/; Wiemann et al. 2001). (B) Artifacts of the corresponding subcellular localization class are shown in the right column: mitochondria artifacts (116 images), endoplasmatic reticulum artifacts (99 images), microtubules artifact (67 images), Golgi artifact (77 images). In total 1035 images were labeled as artifacts and assigned to this class. Note that artifact cells show expression levels both below and above the level of expression typically observed for valid cells (left). Hence, the level of expression is not sufficient to discriminate artifact cells from valid cells (see also Supplemental Fig. 1).

Figure 3

Accuracy of phenotype classification. (A) Classification accuracy determined on the test set for three different classification algorithms using the first 25 ranked features obtained by three different feature selection methods. In contrast to the BayesANN approach, the ANN/GA runs with nonconvergent fivefold cross-validation as regularization. As the result of the BayesANN classifier is dependent on random initialization of network weights, the averaged classification accuracies and the corresponding standard deviations of 10 classification runs of the same test set are shown, while the ANN/GA generates many similar networks during the evolutionary search process. Only the SVM was trained with the entire set of 323 features. (B) While BayesANN performed best on average in combination with SAM for feature selection, the overall best classifier was obtained by combination of BayesANN with STEPWISE (82.6% accuracy). The cone diagram shows single class accuracies obtained by this overall best classifier for all subcellular classes and artifact class.