Predicting cell shapes in melanomas

One of the greatest challenges facing melanoma research is predicting which tumours are the most likely to metastasize and which can be easily cured by surgery. The best indicator of metastatic potential is presently the Breslow thickness, which is simply a measure of how thick the tumour is (Breslow, 1970), combined with a few other features such as ulceration and spread to nearby lymph nodes. Even small or thin melanomas can be aggressive though, and once melanoma spreads, it has a very poor prognosis, perhaps due to the high levels of motility and the plastic nature of melanocytes and melanoma cells. Yin and colleagues recently reported a high-throughput analysis based on cell shape to predict which genes might control metastatic dissemination of melanomas and they report an intriguing siRNA screen identifying several new genes controlling cell shape. Use of computer software to model biological systems or to automate analysis of biological data has become an increasingly attractive concept to researchers and has great clinical potential. Automated analysis helps to remove inconsistencies due to human subjectivity. This is particularly true with the analysis of image data from cell biological research and may be applicable to pathological specimens in the future. A second benefit is the power of computers to process large data sets extremely rapidly. One area of research that is particularly amenable to analysis by computers is cell shape (Bakal et al., 2007; Keren et al., 2008), which may be a predictor of migratory potential and for tumours could indicate metastatic tendencies or other pathological features. While it would be exciting to use automated analysis of human tumour samples to make predictions about their metastatic potential, high-throughput imaging still is generally carried out with cultured cells on rigid substrates. Cells such as goldfish keratocytes can be described by only a few parameters (Keren et al., 2008), but other cell types, such as Drosophila BG-2 cells, which are a central nervous system cell line, require hundreds of different features to describe (Bakal et al., 2007). Mammalian cells in general assume varying complex shapes in culture, which may even be a continuum and which certainly depends on the matrix composition and geometry. Despite the complexities, a recent study has made progress towards analysis of cell shape to help understand the behaviours of human melanomas. The authors used ‘machine learning’ where they trained a computer programme to recognize five distinct human-defined shapes of Drosophila Kc167 cells, a cell line derived from embryonic hemocytes: (i) normal (ii) elongated (iii) small and partially polarized ‘teardrop’ (iv). large and flat with smooth edges and (v) large and flat with ruffled edges. A sixth classification contained cells that had not been placed in any of the five main categories; this accounted for approximately 2% of cultured cells, suggesting that the five categories were mostly sufficient to describe the overall cell shape. For each reference shape, a support vector machine (SVM) classifier was generated such that any given cell could be measured in terms of its ‘similarity’ to that reference shape. Each segmented cell was then categorized by its quantitative morphological signature (QMS), which is a vector containing the cell’s similarity score relative to each of the five reference shapes. The software was then used to analyse the immense amount of image data produced by an siRNA screen (comprising 899 dsRNAs) in Drosophila cells (see Figure 1 for a simple flow diagram). To address how cells transition from one shape to another, a RIFT (rate of intermediate forms or transitions) score was generated, which is a measure of how often the cells exist in intermediate shapes, indicating transition states. Intermediate shapes were rare in Kc167 cells, but were increased with knockdown of two genes, Stam and CamKII. Cells are thus thought to show rapid switch-like interchange between states, a feature that was also found by Yin et al. for melanoma cells plated on collagen gels, but not on rigid 2D substrates. Thus, environmental conditions and cell types can influence whether cells exhibit switch-like or continuous shape changes. It may seem surprising that only five shapes could describe the entire repertoire of normal and siRNA treated cells. This may imply that cells only assume a limited number of shapes in a homogenous environment, such as a culture dish. It may also reflect that the researchers in this case defined the five shapes and taught the computer to recognize cells according to these shapes, so the number of shapes was in effect predefined by the researchers. For their secondary screen in melanoma cells grown on a collagen matrix, the authors focussed on two main shapes: rounded and elongated. Significantly, 14 of the 15 chosen gene hits from the more complex Drosophila screen showed an effect on cell shape in at least one melanoma cell line (Figure 1). This validates the use of large-scale Drosophila cell culture screens to identify potentially interesting melanoma genes and consequently led to the authors predicting a Coverage on: Yin, Z., Sadok, A., Sailem, H., McCarthy, A., Xia, X., Li, F., Garcia, M.A., Evans, L., Barr, A.R., Perrimon, N., Marshall, C.J., Wong, S.T., and Bakal, C. (2013) A screen for morphological complexity identifies regulators of switch-like transitions between discrete cell shapes. Nat. Cell Biol. 15(7), 860– 71.