One group has estimated breast cancer outcome by using machine learning to create a 70 gene prediction algorithm while we and others have used machine learning, in combination with discrete signaling pathways, to predict metastasis-free survival. Others have attempted to distinguish between different types of cancer using many types of algorithms including Support Vector Machines, Principal Component Analysis and Artificial Neural Networks. Yet others predict chemosensitivity on the basis of gene expression and signaling networks. However, while all these approaches have made Ruxolitinib impressive strides and are useful in clinical practice, these ideas have not been combined to produce a principles-based approach to cancer drug design. Here we propose a framework for designing cancer treatments that extends existing ideas using the classifier conceptualization. The first step is to discriminate between cancerous and healthy cells. Because this is the goal of cancer treatment, it is important to concretely specify this goal. To do this, we use the mathematics of classification algorithms in conjunction with measurements of cell markers. The classifier answers the following questions: how much do cancer cells differ from healthy cells, and which biological markers can distinguish them? It asks this question while explicitly considering the heterogeneity of both populations. The markers could include gene expression, surface proteins, etc. Because distinguishing between cancer and healthy cells requires taking into account the heterogeneity in each population, we have focused on markers for single cells rather than cell populations where possible. Classifying algorithms in this context are designed to give the maximal separation of cancer cells from healthy cells in terms of these distinguishing markers. As such, we can say that the results of classification describe what a hypothetical “optimal” drug acting upon these markers could achieve. Part I of the Results demonstrates how one could define this optimality objective using gene expression in single cells and explores how many markers and cells are required to achieve this goal. The next step is to understand how the available treatment tools allow us to utilize the distinguishing markers of cancer cells. We should strive to approximate the optimal drug by designing new drugs or combining existing drugs. Here we focus on the second. Drugs should ideally target distinguishing properties of cancer, but most drugs used in the clinic do not do this perfectly. Furthermore, their mechanisms of action differ. Thus, it may be possible to predict how existing drugs should be combined to produce more desirable results. Actually making this prediction would hinge on relating the drug actions to the distinguishing properties of cancer. For example, if cancer cells differ from healthy cells primarily via three distinct markers.