Deeper Insights into Hematological Oncology Disorders via Single-Cell Phospho-Signaling Analysis

Nobody would doubt that hematological oncologic disorders originate at the single-cell level. As early oncologic events have their origin in epigenetic changes in gene expression or frank mutation at the level of DNA sequence, it has become clear that researchers must drive backwards in our analysis of cancer toward initiating oncogenic single-cell events if we will truly understand and defeat cancer. For such analysis, leukemias and lymphomas provide classic opportunities to move from mouse models to human studies for mechanistic understanding as well as in determination of therapeutic options. First, the normal differentiation pathways are well understood for many of the cell lineages that might eventually be recognized as hematological cancers. Because of this understanding, diversion from normality, at least in terms of gross morphological indicators, has standardly been used to classify hematological tumors according to their various subtypes. Simply by visual confirmation one can often discern the presence of rare abnormal cells in a background of normal peripheral blood or bone marrow cells. Of course, this requires the skills of a trained pathologist, sufficient cells present in the microscopic field to be recognized as diseased, some sort of ‘biomarker’ or indicator of abnormality, and usually backup tests or other symptomatology that will allow the clinician to decide upon a formal diagnosis to provide to the patient. When the clinician prescribes a course of treatment the obvious goal is the elimination of the entire cancer—including all cells that are currently causing manifest symptomatology, as well as killing off those cells that we now think of as cancer stem cells^1,–3 that might remain residual and divide once again to haunt the patient post-therapy.

Interestingly most therapies in cancer treatment are scored and classified, as well as progress through clinical trials, based not necessarily on the complete elimination of the cancer but primarily on partial regression (debulking) of the tumor as read out by lower incidence of cells in the blood for most leukemias, regression of lymph node swellings in lymphomas, or diminished x-ray readings of tumor in multiple myeloma, for instance. The inference is that fewer cancer cells are alive and that the therapy is at least in part targeting the cancer. However, we know that things are more complicated. We understand now that cancers might have at their core eternally self-perpetuating stem cell-like cells that give rise to numerous “daughter” cancer cells of lesser replicative capacity.^1,–3 

Therefore, if we base our treatments on outcomes such as gross debulking of the daughter cancer cells, will we lose sight of therapies (i.e., never having scored their utility correctly in the clinical trials) that might actually kill the stem cells of cancer? Will cancer therapy eventually require the use of two agents to be most effective: requiring therapies that debulk the tumor (the daughter cancer cells) as well as a separate agent that aims to kill the stem cells of cancer? Do all hematological cancers have such rare stem cells, or are there cancers that are themselves nearly entirely composed of such self-perpetuating stem cells? If the stem cells of cancer are in fact a distinct population of cells at the beginning of a pathologically differentiating chain of cell division events that eventually kills the patient, we will of course require techniques that allow us (1) to distinguish such cells both from the bulk cancer itself as well as (2) distinguish them from cells that would otherwise be considered “normal” stem cells. Where are the techniques that allow us to do this today?

Similarly, there are expectations that as the cancer continues to develop and grow in the body, additional heritable changes accumulate in both the stem cells of the cancer as well as the cancer cells themselves. Though not formally proven for every hematological cancer, we might expect that given sufficient time the cancer might develop in a single patient into multiple different clonotypes—each of which might be subtly, or starkly, distinct in the kind of disease it causes as well as the therapies to which it would be susceptible. Certainly we understand that such different clonotypes must preexist in patients prior to therapy, since in many cases genetically and phenotypically similar drug-resistant clones of cancer cells recur in patients. Again, single-cell techniques will be required to understand the numbers of clonotypes a patient has at presentation (to the limit of detection) and during therapy, as well as which of those clonotypes will respond to the therapeutic options the clinician has available. Most strategies to date have depended upon surface markers and flow cytometry to distinguish cancer cells from normal cellular counterparts in the blood or bone marrow. This is been highly effective, but has clear limitations since it assumes that differences between normal cells and cancer clonotypes will be manifest at the cell surface as epitopes that might be recognizable to antibodies or other detection agents that are at hand.

At the most basic level the differences in cancer stem cells, as opposed to normal stem cells—or even in how cancer cells might differ one from the other—are really determined by how such cells process signals from their environment. Environment in this sense means internal cellular events as well as signals received from other cells, extracellular matrix elements, cytokines, etc. As such, one can think of cells as information processing devices: receiving signals, processing them, and making decisions via internal algorithms. Subtle differences in how a given cancer cell processes information—such as response to drug therapy—is really the underlying determinant of whether the therapy will be effective (such as “does the cell have an algorithm that allows it to be induced to apoptose?”). How can we reveal differences between cells when we don’t really know the subtle mutation or epigenetics that would distinguish such a cell?

One approach to get at this nearly intractable problem is to capitulate and realize we might never know all the pathologies of a cancer cell, but we could define all cancer cells of a given class by interrogating those cells with a series of questions and determining their responses in a defined series of outputs. Essentially, we are doing a “psychological profile” of a cell—ask it twenty questions and give it a multiple choice answer sheet to score itself. One would then need a relatively universal set of readouts that were indicative of the information processing state (frankly, cell signaling status) of the cell. Such readouts would best be comprised of signaling nodes previously determined experimentally to be key elements of the cell’s processing network (for instance, MAP kinase proteins, cell cycle phosphorylation events, lipid constituents). The hypothesis would therefore be that with a sufficient number of inputs or interrogations of the cell and a measurement of readouts responsive to these inputs, you could rapidly determine how the inputs correspond to activation of the internal network. If appropriately analyzed, this could also give you the connectivity map between the individual components of the network. In conventional terms this would be considered the signaling map. While it does not provide all the information about the subtle pathologies of the cell, it tells you how such pathologies effect information flow through essential signaling elements you have decided might be critical for regulation of the cancer’s response to the environment. Once you have determined the network response or map you could use this to classify cells in a patient according to disease status or drug response prior to any therapeutic intervention.

A uniquely useful series of markers that are dynamically determined by the environment and universally considered to be indicative of the information processing status of the cell would be phosphorylation of cellular components such as proteins and lipids.^4,–6 Phosphorylation occurs rapidly, controls almost all cellular processes, and is sufficiently stable and distinct as a chemical entity that antibody reagents can be developed to recognize the presence or absence of phosphorylation on proteins or lipids. Using antibodies against phosphorylated proteins researchers have mapped numerous cell-signaling pathologies related to cancer, as well as demonstrated the molecular entities and mutations driving their function. Since the literature is replete with studies of this nature, they will not be reviewed in any manner here. To address the above issues of measuring signaling states at the single-cell level, my laboratory several years ago embarked upon a series of studies wherein we stained cells with fluorophore-conjugated antibodies against phosphorylated proteins and then quantitatively determined the levels of phosphorylation using a flow cytometer.^7,8 With this we were first able to measure the phosphorylation state of as many as 11 proteins or lipids in multiple cell subsets (such as B cells, T cells, or macrophages).⁸ Such an approach allowed for unprecedented access to the intracellular signaling state of primary cells in the immune system.⁸ No longer would one be limited to measuring phosphorylation states by Western blots, wherein the average of the entire population of disparate cells is determined after lysing the cells in question, mixing all of their cellular contents, and thereby losing all of the cell-specific information that we’ve argued above to be so relevant to truly understanding the cancer disease state. By this new approach we could bring the Western blot to the single-cell level, simultaneously measuring the intracellular signaling state while according each signaling state to the relevant subset of cells (e.g., T cells, B cells, macrophages or monocytes etc.). Critically, this allowed for us to access the intracellular signaling states of primary cells right out of the animal model, or most importantly directly out of the blood of a human patient.^9,–13 Taken together with the notion that one can map cellular signaling networks at the single-cell level this allows for patient-specific signaling analysis of the immune system simultaneously with the analysis of the cancer cells themselves.

Using this approach in a series of studies we have previously shown that it is possible to map the signaling status of immune system cells in murine and human primary samples.^9,–16 We started first by demonstrating that it would be possible to quantitatively and accurately measure the phosphorylation status of up to 11 phospho proteins in primary cells.⁸ The technique is relatively straightforward.⁸ One begins with a population of cells in culture that are stimulated with an inducer, such as a cytokine, for a set period of time. The induction event is expected to stimulate a series of signaling changes within the cell (see Figure 1, Color Figures, page 509, for details). Examples of inducers would be cytokines such as interleukin (IL)-2, IL-3, or granulocyte-macrophage colony-stimulating factor (GM-CSF), or proxy activators such as antibodies that cross-link cellular signaling components at the cell surface. Within 10 seconds or up to 15 minutes or an hour (as determined by the experimental design and phospho proteins being studied) a fixative such as paraformaldehyde is applied to the cell mixture with the outcome being that the cell signaling systems are frozen in place by the crosslinking of the fixing agent. Phosphorylation is now stopped, dephosphorylation is largely inhibited, and the researcher can now proceed to permeabilize the cells with alcohols or detergents that will allow access for fluorophore-conjugated antibody staining reagents to the intracellular signaling molecules and the phosphorylation epitopes upon the protein surfaces. Cells are also stained with fluorophore-conjugated antibodies specific to cell surface components (to demarcate T cells, B cell, monocytes, etc.) as well as antibodies to the intracellular phospho protein epitopes. A representation of this is shown in Figure 1 (see Color Figures, page 509. At this stage, cells are now ready for flow cytometric analysis and quantitative determination of the fluorophore levels associated with each antibody, which in turn are bound to the phosphorylation epitopes on various important proteins the researcher has predetermined. Post collection of the data, relatively standard or advanced analysis of the cellular subsets associated with varying signaling profiles can then be carried out (for methods and reviews see ^{7,–9,14,17,–21}.

We followed this technical feat with studies showing that primary CD4-positive T cells from mice required crucial accessory signaling from integrin proteins as initiated by ICAM-2.^13,22 This was demonstrated by single-cell phospho protein analysis to be mediated via a time-dependent early c-Erk signaling event that determined the functional commitment of the T cell. Other studies later showed that similar signaling analyses could be applied to other cellular subsets.¹¹ 

An important advance for us was the demonstration that when we applied this technique to cells from patients with acute myelogenous leukemia (AML), it was possible on the basis of cellular activation states to determine prognostically which patients would respond to chemotherapy (i.e., an initial debulking of the tumor), and those that would not.^12,14 In this study the basal phosphorylation state of six proteins was determined (these proteins were Stat1, Stat3, Stat5, Stat6, p38 and Erk1/2) along with their phosphorylation state as induced by Flt-3, GM-CSF, granulocyte colony-stimulating factor (G-CSF), IL-3, and interferon (IFN)-γ.^12,14 The fold change of induced phosphorylation over background for each AML cell population in each patient was determined. On a patient by patient basis this information was used to cluster those signaling states that were most like another (in a dendrogram) and then it was determined statistically whether any of the groupings correlated with patient response to therapy. It was clearly shown that the basal state was uninformative to predict patient outcome; but when the basal states were used in conjunction with the fold change over background certain signaling signatures could be used to prognosticate therapeutic outcome. As suggested above it was possible to rapidly create patient-specific signaling maps that suggested fundamental signaling differences between those patients who would respond to the chemotherapy versus those who would not.

At one level using phospho-protein signaling analysis as a biomarker is not so different from those studies that have used mRNA chip analysis to determine disease subtypes in patients. The crucial difference in the analyses that we have undertaken is that we measure the response to a signal and accord subtypes via the classification of the signaling connectivities as defined by ‘artificial’ and distinct input states provided in the environment. The unexpected finding was how crucial signal induction was to separating patients according to therapeutic outcome. In retrospect, given the logic outlined above we can now see how forcing a cell to react to its environment is going to allow us to use relatively generic phospho-signaling nodes in conjunction with different environmental stimuli to map the internal signaling network of the cell. Frankly, signal induction, input output measurements, are how biochemists and geneticists have standardly mapped signaling and genetic regulatory pathways for the last 30 to 40 years (an analogous term a friend of mine in the military called it when I explained this to him was “recon by fire”). When one compares this to studies carried out by mRNA chip analysis in which one must sort through up to 30,000 different gene sequences to find a subset that strongly correlate with disease outcomes, the fact that only five signaling inputs and six signaling readouts were sufficient to strongly correlate with therapeutic outcome was demonstrative of the power of using signaling connectivity maps as the drivers of clustering patients most like each other in signaling response.

Finally, as regards this study it was evident that patients with nonresponsive forms of AML had a higher disposition to have more than one subpopulation of cells with discrete phosphorylation signaling states. For instance, one particularly informative intersection of environmental induction coupled to signaling readout was the G-CSF induction of Stat3 and Stat5. In patients who showed an initial positive therapeutic outcome there was little if any change in the co-induced levels of Stat3 and Stat5 (Figure 2, panels ABC; see Color Figures, page 509). However, co-induced signaling in patients who had a negative therapeutic outcome in response to drug showed a massive co-upregulation of Stat3 and Stat5 (Figure 2, panels D, E; see Color Figures, page 509). Thus, it was clear that there was a distinct difference in the signaling architecture in the cells of patients who would respond to chemotherapy versus those that would not. This suggests that epigenetic or genetic changes have occurred that have selected for cells capable of responding to signaling events in the environment. Further, cells with this signaling signature show lack of response to chemotherapy.

However, closer examination of the flow cytometric plots shows something even more revealing. Patient 14 and patient 11 not only have the strongly double-positive Stat3/Stat5 population but they also have clear subpopulations of single positive Stat3 or Stat5 as well as a Stat3/Stat5 low expressing population of cells. Patient 14 and patient 11 similarly have multiple subpopulations of cells. Therefore, each of these patients has evolved what we term multiple distinct “signaling clonotypes”—that is, several populations of cancer cells each with distinct signaling signatures. Perhaps then it is not surprising that the signaling response to standard chemotherapy might differ between these different signaling clonotypes. In addition, more aggressive cancers, as defined by those that did not respond to chemotherapy, had a different signaling architecture than those that did respond—and they had differing ratios of cells in each of the distinct signaling states. Even closer examination reveals that patients 26 and patient 23, who were initially responsive to chemotherapy, both have a small subpopulation of double-positive cells. The question arises as to whether or not these patients are in the initial phase of growing a population of cells that will eventually become the predominant form of the cancer, and simultaneously present as insensitive to the standard chemotherapeutics that were applied in the studies (the patients in this cohort were not followed over time, unfortunately). This same series of flow cytometric plots simultaneously (1) points to a hypothetical clinical progression from patient 27 through patient 11 (left to right in Figure 2; see Color Figures, page 509); (2) suggests a mechanism by which cellular subsets appear that aggressively outgrow their founding population; and (3) begs the question of whether such distinctive subpopulation analysis could be used in vitro to test new drugs that would work against the double-positive cell populations and provide therapeutic options for patients such as 14 and 11.

In studies recently published we have now shown in this same cohort of patients that additional signaling deficits related to the p53 protein could be correlated to extant levels of the anti-apoptotic protein Bcl-2²⁴ and appear to be driven by the mutational status of the FLT-3 receptor in these patients. Finally, signaling defects are also observable in primary patient material other than AML. A 300-patient trial to apply the same style of signaling analysis to follicular lymphoma with the goal of simultaneously measuring the follicular lymphoma cells as well as the infiltrating T cells and B cells has begun.²⁴ Early indications show^15,16 that clearly distinct signaling patterns can be revealed in different patients and mapped to distinct signaling nodes downstream of the B cell receptor. Interestingly, infiltrating T cells and normal B cells showed different patterns of reactivity to the lymphoma microenvironment. Further analysis of the studies will be presented at the 2006 American Society of Hematology meeting.

It is heartening to consider that we now have tools at our disposal to begin to allow us to pick apart distinct cellular subsets in complex tumor tissues. While phospho-protein analysis at the single-cell level by flow cytometry cannot assess the vast number of genes accessible by mRNA chip analysis, the studies we have been undertaking do allow for rapid determination of signaling-responsive states as networks that had, previously, not even been determinable in cell lines. While the discussion focused here entirely on hematological malignancies, studies in our group have also involved other hematological disorders of the immune system such as autoimmunity as modeled by rheumatoid arthritis (RA) or systemic lupus erythematosus (SLE). As with the hematological malignancies the basal phosphorylation states have been relatively insufficient to provide correlations or informative outcomes related to disease state or therapeutic prognosis. However, intelligent selection of internal signaling molecules to be studied (usually those previously associated with the disease state as mined from the literature) have provided fascinating mechanistic insights to the disease while simultaneously providing pharmaco-dynamic monitoring biomarkers (M. Hale, P. Krutzik, GPN for SLE and O. Perez, GPN for RA).

Therefore it is clear there will be multiple opportunities to use and advance the technology in direct analysis of patient samples in clinical trials. Furthermore, as we have shown that there are deeper correlations in the data sets when one is measuring multiple phosphorylation states simultaneously, the data can be used to automatically construct signaling network maps using advanced computational approaches such as Bayesian networks, as we published previously in Science.²³ Combining Bayesian network analysis for automated production of signaling network states with clustering of patient outcomes provides an exciting new opportunity for clinical researchers to move beyond cell lines and generic signaling maps toward a day when it will be possible to provide point-of-care analysis of signaling states along with personalized patient-specific therapies tailored to disease cell subtypes that are truly definitive of the complex disease phenotypes found in hematological malignancies.