High-dimensional single-cell cancer biology

Abstract

Cancer cells are distinguished from each other and from healthy cells by features that drive clonal evolution and therapy resistance. New advances in high-dimensional flow cytometry make it possible to systematically measure mechanisms of tumor initiation, progression, and therapy resistance on millions of cells from human tumors. Here we describe flow cytometry techniques that enable a "single-cell " view of cancer. High-dimensional techniques like mass cytometry enable multiplexed single-cell analysis of cell identity, clinical biomarkers, signaling network phospho-proteins, transcription factors, and functional readouts of proliferation, cell cycle status, and apoptosis. This capability pairs well with a signaling profiles approach that dissects mechanism by systematically perturbing and measuring many nodes in a signaling network. Single-cell approaches enable study of cellular heterogeneity of primary tissues and turn cell subsets into experimental controls or opportunities for new discovery. Rare populations of stem cells or therapy-resistant cancer cells can be identified and compared to other types of cells within the same sample. In the long term, these techniques will enable tracking of minimal residual disease (MRD) and disease progression. By better understanding biological systems that control development and cell-cell interactions in healthy and diseased contexts, we can learn to program cells to become therapeutic agents or target malignant signaling events to specifically kill cancer cells. Single-cell approaches that provide deep insight into cell signaling and fate decisions will be critical to optimizing the next generation of cancer treatments combining targeted approaches and immunotherapy.

Fig. 1

Multidimensional single-cell analysis pinpoints tumor cell signaling. In this example of 10 representative tumor cells analyzed under five stimulation conditions, oncogene expression marks three distinct populations of cells with contrasting signaling responses. In the top row, the number in each cell indicates the level of signaling in that cell under each condition. These values lead to the results shown as “Signaling”. An aggregate analysis might mistakenly be interpreted to suggest that three of the conditions (Stim B, 0.5× Stim A, and Stim A + Drug) elicited the same signaling responses. However, the single-cell view reveals key subset-specific signaling differences. For example, the signal from Stim B is not half as effective as Stim A. Stim B is completely effective at stimulating one subset and ineffective at stimulating another. The oncogene-high cells are hypersensitive to Stim A and non-responsive to Stim B. Similarly, the partial effect of the Drug is due to complete inhibition of one subset and no inhibition of another. Adapted from (Krutzik et al., 2004).

Fig. 2

Abnormal signaling in cancer cell networks. Gains and losses of signaling drive oncogenesis and tumor progression. This figure classifies commonly observed signaling alterations according to direction (potentiated or attenuated) and mechanism. Basal signaling disruptions are commonly observed in cancer cells, and the signaling networks of the most negative prognostic cells typically display altered responses to environmental cues. Refer to (Irish et al., 2006a) for example cancer hallmark signaling changes conferred by gene mutations.

Fig. 3

Discovery and validation of a clinical signaling profile. During the training phase, many hypotheses are tested as signaling is assessed at many nodes under a large number of conditions (basal, various signaling activators, doses, time points, drugs, and combinations). The signaling profile is then refined by determining which features differed in the experimental group (cancer) relative to controls (healthy). This feature selection step is based on the biosignature hypothesis (Irish et al., 2004a), which proposes that features that vary as much in the control group as they do in the experimental group are not likely to productively contribute to unsupervised stratification because they are not specific to the experimental group. Models based on one or more features are then built, and it is determined whether they stratify a feature of interest such as clinical outcome. This clinical signaling profile is then tested in a new set of samples comparable to the first and balanced for potential confounders. Ideally the test is performed by a new investigator or a computer algorithm that is blinded to the outcomes.

Fig. 4

Identifying contrasting signaling in cancer and non-malignant cells of same lineage cell within a tumor. In this example, non-malignant tumor infiltrating lymphocyte (TIL) B cells are detected within follicular lymphoma B cell tumors from two patients. On the left, non-tumor cells were identified by the expression of the “wrong light chain” – a B cell receptor immunoglobulin light chain of a different isotype from the clonal tumor – combined with high CD20 expression and a lack of BCL2 expression. Here we can see that these cells have a distinct SYK and BTK signaling profile that contrasts with the bulk tumor. The histogram overlays on the right show potentiated magnitude and kinetics of ERK and p38 phosphorylation in lymphoma B cells (right side, BCL2⁺) vs. TIL B cells (left side, identified as λ⁺ non-tumor light chain and BCL2^-).

Fig. 5

Hypersensitivity to a signaling input is diagnostic for JMML. a Previously, 3-4 weeks were required to confirm a suspected diagnosis of JMML with a granulocyte-macrophage colony-forming units (CFU-GM) assay. In the CFU assay, bone marrow cells from healthy donors (green curve) and patients have different responses to GM-CSF. b Plot of colony growth vs. GM-CSF dose in healthy volunteers (green) and patients (red).c By flow cytometry, a hypersensitive population of JMML cells is detected in cancerous bone marrow compared to the normal control. d A dose-dependent increase in hypersensitive activity of p-STAT5 uniquely distinguished JMML from other myeloproliferative disorders as well as healthy patients. Adapted from (Kotecha et al., 2008).

Fig. 6

Emergence of a negative prognostic subset over time following treatment. In this example, LNP tumor cells from lymphoma patient J038 are distinguished by abnormal SYK and PLCγ signaling and differential BCL2 and CD20 expression (gold arrow). At the time of diagnosis, LNP cells constituted only 46.3% of the tumor cells. After therapy and disease progression, LNP cells increased to 68% of the tumor. Each 1% increase in LNP cells is associated with a 2.5% increased risk of death in the following year (p < 0.000005, z-score = 4.68). Adapted from (Irish et al., 2010).

Fig. 7

A hallmark mechanism of AML therapy resistance is rewired JAK/STAT signaling. In this example, signaling profiles of two different AML cancer cells are shown. In treatable AML cells, G-CSF signaling through JAK1 and induction of STAT5 phosphorylation mediates transcription of pro-survival and proliferation genes. Conversely, IFNγ signaling through JAK2 results in induction of STAT1 phosphorylation that mediates cell cycle arrest and apoptosis. In the signaling network of the therapy resistant AML cell, the response to IFNγ has become rerouted to STAT5, which, like G-CSF, mediates transcription of pro-survival and proliferation genes. The lack of functional STAT1 activation, which activates cell cycle arrest induced apoptosis, explains why patients with these cancer cells are often resistant to DNA-damage-induction therapy. Inhibition of JAK2/STAT5 signaling in therapy resistant AML cells could potentially improve the outcome of patients with this resistant subset.

Fig. 8

Key single-cell opportunities in cancer research. The first row depicts the opportunities of detecting non-malignant cells of the same lineage as the tumor (A, as in Fig. 5), tumor infiltrating immune responders (B, as in (Myklebust et al., 2013)), and other non-malignant stromal cells (C). It will be important to distinguish between abnormal signaling that promotes cancer, such as inflammation, and abnormal signaling that results from cancer, such as T cell suppression via PD-1 or generation of cancer associated fibroblasts (Barcellos-Hoff et al., 2013). In the second row, (D), (E), and (F) depict contrasting biological origins of an aggressive, therapy-insensitive tumor subpopulation that can be dissected with single-cell tools. A gatekeeper mutation conferring resistance to targeted therapy might be an apomorphy that distinguishes a rare ‘leaf’ subset (F). Alternatively, a slow cell cycle phenotype might distinguish a cancer stem cell (D)(Reya et al., 2001). A large, heterogeneous branch (E) observed at the time of diagnosis might need to be treated with a combination of therapies in order to kill all populations and obtain a clinical response. The third row depicts clinical single-cell opportunities, such as detecting negative prognostic subpopulations (G, as in Fig. 7), treatment insensitive subsets (H), and cellular transitions as would be observed when epithelial cancer cells become an invasive, metastatic population (I).