If cancer drugs behave as classifiers for colon cancers, they will likely behave as classifiers in other organs. Analyzing cells originating from different tissues is therefore not likely to affect the principle of this study. Another limitation of this study is that we VE-821 1232410-49-9 measured markers in different tissue contexts. Unlike the Quake dataset, the Gray dataset measured expression of cultured cell lines rather than expression levels in primary single cells. Additionally, because this dataset did not include healthy cells, we had to demonstrate our approach by targeting one subtype of breast cancer versus another, rather than cancerous versus healthy cells. More rigorous analyses based upon treatment of single-cell or minimally amplified clones as well as relevant classification analyses will have to wait for more extensive experimental data to become available. An important question that this study highlights is whether experiments with tumor cell populations capture as much relevant information as experiments with individual cells. Cell line gene expression levels reflect a population average rather than the expression of individual cells. It is possible that small numbers of drug-resistant cells with distinct marker profiles may not be detectable when the whole tumor cell population is analyzed together. For example, small populations of cells within a tumor have been shown to drive the evolution and drug resistance of some types of cancer. But although single cell data is preferable, it presents challenges of its own. Even genetically identical single cells can differ phenotypically in a number of ways including epigenetic state, protein and gene expression and morphologic state, and these states can change with time. Thus, nongenetic heterogeneity across single cells may complicate classification analyses. These factors underscore the need to measure a diverse set of markers from single cells in order to target cancer effectively. Our statistical analyses have some similarity to other cancer studies that used statistical approaches in conjunction with large datasets. The Friend group, for example, used machine learning to predict prognosis and phenotypes in breast cancer. This resembles our strategy in that they use gene expression to make predictions about cancer. Other groups have used machine learning to distinguish between sub-types of cancer. They typically combine microarray data with such algorithms to distinguish between types of cancer. This resembles our strategy to optimally discriminate between cancer sub-types. The Golub group used machine learning to classify cell lines as belonging to a drug-resistant or drug-sensitive class based upon gene expression whereas the Gray group related certain signaling networks to drug response. More recently, a large consortium predicted cancer drug sensitivity using gene expression. These studies have a similar goal of modeling drug responses.