New Perspectives on Leukemia Initiating Cell Dynamics - Redefining Concepts By Tracing Single Tumor Cell Lineages Through Barcoding Technology

1 1. Introduction 2 3 1.1 Cancer and Clonal evolution Cancer is sustained by production of aberrant cells that vary in many morphological and physiological properties. The resulting intratumor diversity remains a major challenge to our understanding of the neoplastic process and development of therapeutic resistance (Kreso et al. 2013). In 1971, Knudson published the 2-hit hypothesis for mutation-driven cancer based on statistical analysis of inherited and sporadic cases of retinoblastoma. He postulated that retinoblastoma developed as a consequence of two co-operating mutations; one of which could be inherited or somatic followed by a second somatic mutation (Knudson 1971). Cairns hypothesized a different, but complementary, mechanism of tumor suppression in 1975 (Cairns 1975). It was based on that the tissue architecture protected against selection of variant somatic cells with increased fitness in proliferating epithelial populations, such as the intestine and other epithelial organs. He postulated that this could be accomplished by restricting the number of stem cells, for example at the base of intestinal crypts, which restrained the opportunities for competition between cells by shedding differentiated intestinal cells into the gut. In 1976, Nowell first described cancer as an evolutionary process with parallels to Darwinian natural selection. He synthesized the evolutionary view of cancer as a process of genetic instability and natural selection. Most of the genomic alterations that occur are deleterious for the cell, and most clones will thus go extinct, but occasional selectively advantageous mutations arise that lead to clonal expansions. This theory predicts a unique genetic composition in each neoplasm due to the random process of mutations, genetic polymorphisms in the human population, and differences in the selection pressures of each tumors microenvironment. Importantly, the theory also predicts the emergence of resistant clones under the selective pressures of therapy (Nowell 1976). After decades of research, it is now consensually accepted that most neoplasms arise from a single cell of origin (the founding cancer clone), whose descendants sequentially acquire ever more genetic variability during tumor progression (gradually defining additional clones of the same cancer) allowing constant selection of more and more aggressive sublines (Figure 1). Each additional clone in the cancer is defined by distinct mutations that occur during disease evolution, established in some cells but not others. It is increasingly clear that many advanced tumors follow a branched Darwinian evolutionary trajectory, as has been demonstrated in e.g. childhood ALL (Anderson et al. 2011), pancreatic cancer (Campbell et 4 al. 2010; Yachida et al. 2010), colorectal cancer (Thirlwell et al. 2010), clear cell renal carcinoma (Gerlinger et al. 2012; Gerlinger et al. 2014), breast cancer (Navin et al. 2011) and prostate cancer (Haffner et al. 2013). Following branched evolution, multiple cancer clones can co-exist, spatially separated within the same tumor or intermixed within the same biopsy. The presence of multiple clones and the resulting intra-tumor heterogeneity is, however, not synonymous with branched evolution; linear cancer evolution with incomplete selective sweeps may also result in clonal intermixing (Figure 1) (Burrell & Swanton 2014; Swanton 2014; Meacham & Morrison 2013). Figure 1 - Clonal evolution in primary tumors and metastases. The gray circle represents a normal cell, and the central dot depicts the initiating somatic mutation that drives the founder clone in the tumor. The different colored circles represent novel clones of the same cancer that have accumulated successive mutations. Note that in the primary tumor, several different clones co-exist and although some expand, others remain dormant or become extinct. Metastases can originate from either a major clone in the primary tumor (metastasis 1), or from minor clones (metastasis 2). Metastases can also undergo clonal evolution (as shown in metastasis 1). Within each cancer clone there is also significant sub-clonal heterogeneity, with (non-genetic) variability represented by the separation of the different circles of the same color by the black lines. Adapted from (Caldas 2012). Normal cell Founding cancer clone Additional cancer clones with individual genotypes 5 Many studies have further shown that different genetic clones from the same tumor often vary with respect to metastatic potential, driving of recurrence and growth in xenograft assays. Resistance to cancer therapies can indeed arise through discrete genetic mutations found only in certain clones. The observations substantiate the widely accepted view that tumors comprise genetically diverse clones, some of which survive therapy and contribute to disease recurrence. Tumor cell populations appear more genetically unstable than normal cells, perhaps due to activation of specific gene loci in the neoplasm, continued presence of carcinogen, or even nutritional deficiencies within the tumor. Somatic mutations found in cancer genomes may be the consequence of the intrinsic slight infidelity of the DNA replication machinery, exogenous or endogenous mutagen exposures, enzymatic modification of DNA, or defective DNA repair. In some cancer types, a substantial proportion of somatic mutations are known to be generated by exposures. Tobacco smoking in lung cancers and ultraviolet light in skin cancers are classic examples, in parallel with abnormalities of DNA maintenance, for example defective DNA mismatch repair in some colorectal cancers (Alexandrov et al. 2013; Swanton 2012; Kreso et al. 2013; Meacham & Morrison 2013). Our understanding of the mutational processes that cause somatic mutations in most cancer classes is remarkably limited. Since clones often have many parallel genetic and epigenetic alterations in their genomes, it is often not clear which of those alterations that really provide a growth advantage or cause survival benefits and which alterations are simply hitchhiking passenger mutations. Thus, a mutation that gives a cell a fitness advantage is termed a driver and one that has no effect on its fitness and/or growth characteristics is called a passenger. However, this binary classification of cancer mutations into drivers and passengers is context dependent. Tumor (sub)clones compete with each other and with normal cells for ‘real estate’ and resources within the tissue microenvironment. Changes imposed on this ecosystem will alter the relative competitiveness of cancer cell (sub)clones. Minor (sub)clones that prior to therapy were relatively unfit, but then were able to survive an anti-cancer treatment, can afterwards be the drivers of the malignancy. In fact, mutations that confer drug resistance might actually be disadvantageous in the absence of treatment (Grove & Vassiliou 2014; Gerlinger et al. 2012). Cancers generally arise from a clonal origin marked by early founder mutations and/ or DNA copy number alterations. Recent studies employing next generation sequencing (NGS) of tumor genomes have clearly demonstrated most tumors, across many types of cancer, have evolved to genetically heterogeneous entities, harboring multiple clonal populations of 6 cells (Burrell & Swanton 2014; Swanton 2014; Meacham & Morrison 2013). Specifically, whole exome sequencing (WES) of multiple regions within primary tumors has revealed extensive regional heterogeneity in pancreatic, hepatocellular, and renal carcinomas. Moreover, whole genome sequencing (WGS) of sequential tumor samples from patients with de novo acute myeloid leukemia (AML) has demonstrated genetic evolution between diagnosis and relapse and WGS of paired primary-metastasis samples in breast cancer demonstrated the importance of clonal evolution also in this context. Thus, clonal heterogeneity within tumors represents a fundamental drive of the biological complexity of human cancers and a detailed understanding of it is important for better understanding of clinical genomics. The ultimate genetic resolution will be achieved by multi-sampling and single-cell analysis, which is rapidly becoming tractable (Hughes et al. 2014; Shah et al. 2009; Anderson et al. 2011). This notwithstanding, a large part of the intratumor heterogeneity stem from the variation between different cells within each of the genetically homogenous cancer clones, i.e. the sub-clonal heterogeneity (Figure 1). For many neoplasms only a minority of the cancer cells are able to sustain tumor growth, although such cells can comprise the majority of tumor cells for some cancer types. Whether these cells are part of major, or represent distinct, genetic clones remains to be determined. Addressing these questions requires genetic analyses combined with functional assays that measure tumor propagation at the resolution of individual sub-clones derived from single cells. The sub-clonal heterogeneity reflect variability due to differences in e.g. epi-genetic and transcriptomic status of genetically identical cells (Kreso et al. 2013; Meacham & Morrison 2013). The importance of sub-clonal heterogeneity was demonstrated by John Dick’s team in a study published in 2012. By combining DNA copy number alteration (CNA) profiling, sequencing, and lentiviral lineage tracking, they followed the repopulation dynamics of 150 single lentivirus-marked cell lineages from ten primary human colorectal cancers through serial xenograft passages in mice. CNA and mutational analysis distinguished individual genetic clones, and showed that these remained stable on serial transplantation. Despite this, the proliferation, persistence, and chemotherapy tolerance of individual lentivirally marked lineages within each clone were very variable. Chemotherapy promoted dominance of previously minor or dormant lineages. This clearly demonstrated that there apart from the genetic diversity exist another important layer of variation between the cancer cells within each tumor, resulting in additional functional variability in tumor propagation potential, a mechanism that contributes both to cancer growth and therapy tolerance (Kreso et al. 2013). 7 It is urgent to define the impact of intratumor heterogeneity on drug response, biomarker validation and clinical outcome in prospective clinical trials. What is the impact of intratumor heterogeneity on disease outcome and how do cancer selection pressures, both microenvironment and treatment related, modulate cancer evolutionary trajectories? One needs to understand whether targeting a dominant driver of the founding clone results in improved progression-free survival outcomes compared with targeting the same driver when limited to only certain other clones, present at specific stages in the tumor branches. One also needs to determine small clones harboring high-risk driver events that might themselves be targetable to limit tumor progression (Swanton 2014; Ng et al. 2014). What Nowell postulated almost 40 years ago is finally being experimentally confirmed, but we still have a long way to go to until we can utilize this knowledge to improve patient outcomes and his outline still holds: “More research should be directed toward understanding and controlling the evolutionary process in tumors before it reaches the late stage usually seen in clinical cancer” (Nowell 1976). 1.2 Acute myeloid leukemia AML is a heterogeneous and aggressive hematological malignancy characterized by a block in myeloid differentiation and uncontrolled proliferation of abnormal myeloid progenitors that accumulate in the bone marrow and peripheral blood, with possible spread to liver and spleen. Some cases develop from other hematopoietic disorders or follow genotoxic therapy for unrelated malignancies, but most arise de novo. Several genetic markers have been identified to stratify patients into prognostic groups, which are used to guide treatment decisions. Although chemotherapy results in high rates of remission, the majority of patients relapse and the overall 5 year survival is only 40–45% in young patients and less than 10% in the adults (Grove & Vassiliou 2014). GLOBOCAN estimated the worldwide total incidence of AML for 2012 to be 351,965 with an age-standardized rate per 100,000 individuals of 4.7, a 5-year prevalence of 1.5% and a M:F ratio of ~1.4. Unfortunately, the International Agency for Research on Cancer does not sub-classify leukemia into acute and chronic, and myeloid or lymphoid in its GLOBOCAN analysis (WHO 2014). Two systems have been used to classify AML – the earlier French-American-British (FAB) classification and the more recent World Health Organization (WHO) classification. 8 FAB classification - In the 1970-80s, French, American, and British leukemia experts divided AML into 8 subtypes, based on the type of cell from which the leukemia developed and how mature the cells were: ∙ M0 Undifferentiated ∙ M1 AML without maturation (poorly differentiated) ∙ M2 AML with maturation (more differentiated) ∙ M3 Acute promyelocytic leukemia (APML) ∙ M4 Acute myelomonocytic leukemia (AMML) Subtype: − M4 eos : Acute myelomonocytic leukemia with >5% eosinophils ∙ M5 Acute monocytic leukemia Subtypes: − M5a :Acute monoblastic leukemia - poorly differentiated − M5b: Acute monocytic leukemia – more differentiated ∙ M6 Acute erythroblastic leukemia ∙ M7 Acute megakaryoblastic leukemia WHO classification - In 2001, the WHO classification reflected a paradigm shift from previous schemes in that genetic information was incorporated with morphologic, cytochemical, immunophenotypic, and clinical information into diagnostic algorithms for the myeloid neoplasms (Jaffe ES, Harris NL, Stein H 2001). The WHO classification was revised in 2008 to incorporate new scientific and clinical information to refine diagnostic criteria for previously described neoplasms and to introduce newly recognized disease entities. It is a consensus classification in which a number of experts have agreed on the classification and the diagnostic criteria used for defining the entities that compose it. This classification can be found in the 4rd edition of Classification of Tumors of the Hematopoietic and Lymphoid Tissues as part of of the series, WHO Classification of Tumors (Vardiman JW, Thiele J 2015). Patients with AML usually present because of complications of disordered hematopoiesis: bleeding, fatigue, refractory infections; or the clinical consequences of an extremely high white blood cell count: difficulty breathing, confusion, or other symptoms of organ failure. Beyond supportive care to stabilize patients, definitive treatment is typically divided into two phases. Induction therapy, the initial therapy meant to induce a disease remission, consisting of DNA-damaging agents (most commonly cytarabine, a nucleoside analogue, combined with 9 an anthracycline such as daunorubicin), can frequently cause apoptosis in the majority of the leukemic cells, and most patients achieve a remission. Unfortunately, disease will relapse in almost all patients unless it is eliminated by additional therapy. Further treatment attempting to consolidate remission into cure, “consolidation therapy”, consists of additional chemotherapy (cytarabine) or bone marrow stem cell transplantation (SCT) (Showel & Levis 2014; WHO 2014). In patients younger than age 65, although induction chemotherapy can generally yield remission rates of 70% to 80%, at least half of these patients eventually relapse and ultimately die without SCT. Even after SCT, over one third will relapse. Despite the current dismal median survival for AML of approximately one year, overall survival for younger patients with AML has been rising steadily. To some degree, this improvement can be attributed to advances in supportive care (Showel & Levis 2014). The molecular pathogenesis of AML has not yet been completely defined, however there is prognostic risk group stratification for AML based on cytogenetics and novel molecular parameters (Figure 2). Recurrent chromosomal structural variations (e.g., t(15;17), t(8;21), inv(16), t(9;21), t(9;11), del5, del7, etc.) are established diagnostic and prognostic markers, suggesting that acquired genetic abnormalities play an essential role in leukemogenesis. However, nearly 50% of AML cases have a normal karyotype, and many of these lack recurrent structural abnormalities, even with high density comparative genomic hybridization or single nucleotide polymorphisms arrays. However, for AML, normal karyotypes are common, and genomic instability is unusual (Welch et al. 2012). Gilliland and Griffin proposed the two-hit model of leukemogenesis. In this model, two lesions, each belonging to a different class, collaborate to cause AML when neither is sufficient to do so in isolation. Class I mutations such as FLT3-ITD or N-RAS mutations confer a proliferative advantage, but have no effect on differentiation. Class II mutations (represented by specific fusion genes in the original model) impair hematopoietic differentiation and subsequent apoptosis. The initiating lesions in these AMLs are thought to be class II mutations, for example PML/RARα and MLL gene fusions, whereas class I mutations are typically later events. This model has provided a useful framework to conceptualize the pathogenesis of AML as a disease in which differentiation is blocked and proliferation is increased. Although most of the recently identified mutations do not fit neatly into one of the two classes, they are thought to synergistically produce the equivalent effects (Grove & Vassiliou 2014; Gilliland & Griffin 2002). 16 studies showed that these mice allowed for increased human hematopoietic engraftment when compared to SCID mice. The mutation responsible for this favorable behavior was later identified to be in the SIRPA gene, which encodes the transmembrane receptor SIRP on phagocytic cells such as macrophages and dendritic cells. In 2004, a new immunodeficient mouse model was developed by combining recombinase activating gene-2 (Rag2) and common cytokine receptor gamma chain (c) mutations. The Rag2/c-double deficient mice were completely alymphoid, showed no spontaneous tumor formation, and exhibited normal hematopoietic parameters. Interestingly, human cord blood cell engraftment in Rag2-/cmice was greatly enhanced by the exogenous administration of human cytokines, interleukin-3, granulocyte-macrophage colonystimulating factor, and erythropoietin, in contrast to the NOD-SCID model. This unique feature of the Rag2-/cmouse model has been particularly well suited for assessing the role of different cytokines in human lymphopoiesis and stem/progenitor cell function in vivo (Mazurier et al. 1999). While NOD-SCID, Rag1 and Rag2 mutations prevent the recombination of genes required for functional B and T cell receptors, the strains differ greatly in their radiosensitivity. The Prkdc gene disrupted by the SCID mutation is expressed broadly and is involved in DNA repair, while expression of Rag1 and Rag2 is limited to hematopoietic cells and involved only in DNA recombination of T and B cell receptor genes. Thus, SCID mice are significantly more sensitive to radiation-induced DNA damage than their Rag1 or Rag2 mutant counter parts (Pearson et al. 2008). As mentioned above, a major breakthrough was the generation of mouse models with reduced or nearly absent natural killer cell activity. Most commonly used today are mice with deletions or truncations of the common chain (IL2Rg) in a NOD-SCID background. NOD/ShiLtSz-scid/IL2Rg-/- (NSG) mice are a popular example of a mouse model employing this strategy to eliminate natural killer cell activity (Shultz et al. 2005). In 2012, Vargaftig et al. showed that NSG mice are significantly more permissive to leukemic engraftment than the previously used NOD-SCID mice. When assayed in the NSG model the apparent number of AML L-IC increases on average 31-fold compared to the NODSCID model. However, it is important to note that AML samples that did not engraft in NODSCID mice were also not engraftable in NSG mice (Vargaftig et al. 2012). Later, the targeting of the IL2Rg gene in the NOD-SCID strain by either complete knockout or truncation of the intracellular portion (NOD/LtSz-SCID IL2RG: NSG and NOD/shiSCID IL2RG, respectively) was shown to diminish innate immunity, in particular macrophage 17 function and natural killer cells activity, thereby improving engraftment by normal human CD34+ cells and AML cells. Alternatively, transgenic expression of three nonor poorly crossreacting cytokines in the NOD-SCID background NOD/LtSz-SCID-SGM3 (NSS) led to improvements in expansion of normal human myeloid cells, as well as slight gains in engraftment of AML samples (Ito et al. 2002; Shultz et al. 2005). Table 2 – Principal characteristic and improvements of mice models used to study AML over years. Mice Model Year Principal Characteristics SCID 1983 Lacking in T and B lymphocytes and immunoglobulins. No serial transplantation, poor multilineage engraftment, high cell doses required for transplantation. NOD-SCID 1995 Not imuno-specific. Ca not handle much irradiation or chemotherapy. Increased human hematopoietic engraftment when compared to Scid mice. Rag2-/-/c-/- 2004 Completely alymphoid (T-, B-, NK-), showed no spontaneous tumor formation, and exhibited normal hematopoietic parameters. Less radiosensitive. NSG 2005 Absent NK-cell activity. Improvements in expansion of normal human myeloid cells. Slight increase in human hematopoietic engraftment. NSG-S 2010 Improved engraftment of pre-leukemic myeloid cell cultures, as well as primary human AML. In 2010, Wunderlich et al generated a new mouse strain by crossing the NSG mouse with the NSS mouse to generate the NOD/LtSz-SCID IL2RG–SGM3 mouse (NSGS). They saw that the reduced innate immune response increases initial homing efficiency to the bone marrow and may support subsequent growth/survival of the cells. In contrast, human cytokines do not affected homing efficiency but provided cell proliferation and survival signals. These effects seem additive, as evidenced by the superior engraftment observed in NSGS mice relative to NSG or NSS mice. NSGS mice reproducibly showed significantly higher levels of human CD45+/CD33+ myeloid cells compared with NSG mice on pre-leukemic myeloid samples. Thus, NSGS mice offer the opportunity to perform xenotransplant studies for these pre-leukemic myeloid cells. A secondary transplant into NSGS mice resulted in a similar phenotype of the AML cells including the primitive cell population, indicating retention of the leukemia initiating cells in the NSGS mouse. This study suggested that the NSGS 18 mouse is a significantly better host for at least a subset of AML samples relative to NSG mice, the current standard. The combination of human cytokine expression and IL2RG knockout in NOD-SCID strains results in enhanced engraftment of pre-leukemic myeloid cell cultures, as well as primary human AML samples (Wunderlich et al. 2010). 19 2. Aims 20 21 2. Aims of this thesis Cellular lineage tracing technologies are important research tools that make us uniquely able to follow in parallel the fate of all descendants of multiple labeled cells both in vitro and in vivo. In the intersection between the fields of clonal evolution in cancer and L-IC studies, these technologies specifically help us to pinpoint which sub-clones are responsible for phAML establishment or relapse (following therapeutic targeting) in mice. In this thesis I aimed to contribute to the current understanding of L-ICs by studying the in vivo dynamics of multiple different tumor populations, each stemming from a single-celllineage. Specifically, we used the extraordinary resolution of cellular barcoding to elucidate the processes governing disease engraftment, by dissecting the architecture of the tumors establishing in our novel murine xenotransplantation model. We employed a powerful lentiviral-based cellular barcoding technology that allows large scale parallel lineage tracing, by incorporating unique molecular DNA tags (traceable noncoding sequences) in hundreds of individual human AML cells that establish in a new xenotransplantation model of AML, based on intra-bone marrow (IBM) tumor inoculation in immunodeficient mice (Rag2-/-γc-/- for cell lines; and NSG for phAML), established specifically for this project. The specific aims of this thesis were to: 1. Optimize the pre-existing protocols for lentiviral-based barcode transduction, in order to meet the required quality controls for using this novel approach to trace dynamics of single AML cell lineages in our xenograft model. 2. Study the L-IC frequency in a hAML cell line (HEL) based in vivo model and quantify how many of these cells are needed to generate disease establishment. 3. Compare the results obtained by ELDA and lentiviral-based cellular barcoding technology, as two different alternative methods to quantify L-ICs in the very same mice/ xenografted hAML tumors, and determine their relative sensitivity. 4. Study the impact of the microenvironment/niche availability on hAML engraftment/L-IC frequency. 5. To improve the potential translation of our xenograft studies by searching for an alternative hAML cell line that (better than HEL) represents a clinically relevant phenotype in terms of expression of stem cell markers. 6. Provide proof-of-concept data on the use of barcoded primary AML patient tumor samples in an optimized variant of our xenograft model. 22 23 3. Materials and Methods 24 25 Cell culture. Human acute myeloid leukemia cell line HEL (Martin & Papayannopoulou 1982) (derived from human erythroleukemia) was obtained from American Type Culture Collection (ATCC, TIB-180™). HEL cells were cultured in DMEM (1x) + high glucose+ GlutaMAX™ Supplement (Gibco®, NY, USA) supplemented with 10% (v/v) fetal bovine serum (FBS, gold heat inactivated, PAA, Piscataway, NJ, USA) and 1% (v/v) penicillinstreptomycin (PenStrep, 10,000 units/mL of penicillin and 10,000 μg/mL of streptomycin., Gibco®, NY, USA) and incubated at 37ºC in humidified air with 5% CO2. Cells were maintained by replacement of fresh medium every 3 days. Start cultures at 3 x 105 cells/mL and maintain between 3 x 105 and 7 x 105 cells/mL. The same culture conditions were used for GDM-1 (stablished from the peripheral blood of a patient with a Philadelphia chromosome negative myeloproliferative disorder, after transformation to acute myelomonoblastic leukemia (Ben-Bassat et al. 1982) and was obtained from ATCC Collection (ATCC, CRL.2627™) and KG-1( derived from a marrow aspirate obtained from a 59-year-old Caucasian male with erythroleukemia that evolved into acute myelogenous leukemia (Koeffler & Golde 1978) and was obtained from ATCC Collection (ATCC, CRL.246™) human acute myeloid leukemia cell lines. 293HEK (ATCC® CRL-1573™) cells were plated (70.000cells/cm2) in tissue culture flask (in 0.2ml medium /cm2) in DMEM (1x) + high glucose+ GlutaMAX™ Supplement (Gibco®, NY, USA) supplemented with 10% (v/v) fetal bovine serum (FBS, gold heat inactivated, PAA, Piscataway, NJ, USA) and splitted in the same conditions when 90% confluent. Primary human AML patient blasts came from BM samples collected at diagnosis, from patients admitted to IPO-Lisboa. MNC fraction was isolate the by ficoll before freezing. When needed, were quickly thawed, resuspended in 40 mL of IMDM with 20% FBS and centrifuged at 600g for 5 minutes. phAML were cultured in IMDM (1x) (Gibco®, NY, USA ) supplemented with 10% FBS, 50μM β-mercaptoethanol, gentamicin, and human cytokines (as described by (Pabst et al. 2014)). Fresh growth media was added weekly. 32 33 4.1 Establishment of a novel approach to study dynamics of L-IClineages in AML. Analyses of L-IC have primarily relied on LDA with binary (- vs. +) readout of tumor engraftment in each xenografted animal (most notably immunodeficient mice models), with ELDA representing the preferred tool to quantify L-IC frequency. Since based on statistical regression analyses, such systems always rely on the test of multiple replicates at each of several sequentially decreasing cell doses, to enable measurement of the proportions - with dilutions continuing down to (or at least close to) where the proportion hits 0 - that develop leukemia at each (of several) dilution step(s). Regrettably, current methodology also have some more important limitations, most notably that ELDA does not provide the resolution to dissect the conditions where multiple parallel LIC-lineages co-exist. This central caveat motivated us to develop a highly innovative experimental system allowing us to directly quantify several >100 co-existing single-celllineages, each stemming from a distinct L-IC, while collectively composing the entire tumor. We are confident this will bring new key insights into the cancer subpopulation dynamics that underpin tumor establishment, growth and spread. In our laboratory we are using a sophisticated “cellular barcoding” technology, allowing quantitative large-scale parallel longitudinal lineage tracing, within novel therapeutic xenograft models of human AML that we have recently developed. In this thesis, I first employed this model to interrogate L-IC-lineage dynamics during xenograft establishment of humam AML (hAML) in vivo. The cellular barcoding technology was originally developed by our collaborator, Dr. Ton Schumacher (NKI, The Neherlands), to study T cell differentiation (Schepers et al. 2008). The basis is tagging of individual cells with specific molecular “barcodes” (BCs, traceable noncoding DNA segments), which are stably incorporated in the genome by lentiviral vectors and inherited by the daughter cells. These barcodes thus allow quantitative (NGS-based) tracing of the populations/ lineage descending from each tagged cell over time and space. We first generated a library of lentivirus, with each particle encoding one of ~2600 unique barcodes (BC) linked to the cDNA of gfp (LV-BC). Upon transduction, the barcoded cells are thus tagged by GFP expression, giving us the ability to quantify and follow the transduced (Td) cell population. After transplantation and systemic establishment of the barcoded cells in mice, we re-isolate them by FACS. To determine the relative frequency of the different single-cell-derived-lineages, we quantify the barcodes of each population by 34 deep sequencing. The outline of the basic steps to implement the cellular barcoding technology is schematically described in Figure 4. Figure 4 – The basic steps of implementing the cellular barcoding technology in a general in vivo system, with specific tasks numbered from 1 to 5. To implement this technology in studies of L-IC, we transduced a hAML cell line (HEL) with the LV-BC. Following in vitro culture for 19-24h after the transduction, we sorted GFP+ cells by FACS and injected them in titration into the bone marrow (BM) in our xenograft mouse model. This setup provides data for both ELDA and barcode analyses in parallel. Essential technical aspects regarding the transduction, such as the low transduction efficiency and the tracing of a maximum of 10% of the BCs in library, outlined in Figure 4, will be discussed in detail in section 4.1.2. Once the disease establish we followed the Barcode : unique noncoding DNA sequence GFP tag 1. Lentivirus generation Mixture of lentiviral vectors that each encode one of ~2600 unique barcode and GFP 2. Transduction of cells with barcode library: - Low Td efficiency (<10%) - Tracing max 10% of BCs in library Population 1 3. Transplantation of barcoded cells into mice and in vivo selection 4. Isolation of different cell populations Population 2 5. Recovery, and quantitative deep sequencing of barcodes in each population, e.g.: 0 10 20 30 BC1 BC2 BC3 BC4 BC5 % Barcode frequency in Population 1 0 20 40 60 80 BC1 BC2 BC3 BC4 BC5 % Barcode frequency in Population 2 35 progression, by measuring the number of GFP+ cells in the peripheral blood (PB), of xenografted mice over time and at terminal disease sort GFP+ cells from the BM, PB and peripheral organs to analyze the genome integrated barcodes. Specifically, the barcode segments were first amplified by a nested PCR and then sequenced by NGS, to reveal the BC-architecture of each population. We were thereby able to trace how many single-cancercell-derived lineages that exist in each of our xenografts samples and thus how many L-ICs that contributed to the established disease in each animal. The steps to implement the cellular barcoding technology in our hAML xenograft model are schematically described in Figure 5. Figure 5 – Overview of our xenograft model, with representation of how the cellular barcoding technology was implemented. Tasks are numbered from 1 to 6. Sorted barcoded GFP+ HEL cell line bc bc bc bc bc bc bc bc bc bc bc bc bc bc bc bc7 bc89 bc42 bc15 bc33 bc3 bc19 bc68 bc20 bc8 bc31 bc85 bc8 bc61 bc42 bc68 bc14 1. Lentiviral Production and HEL cell line transduction 2. Sorting GFP+ hAML blasts by FACS, after 19-24h of in vitro culture following LV-BC transduction 3. Intra-bone marrow xenograft transplantation of 100/1000/10000/100000 barcoded hAML cells 4. Monitor disease progression by measuring number of GFP+ cells in PB by quantitative flow cytometry Lentivius encoding barcode library 5. Upon disease establishment, collect BM, PB and peripheral organs to sort barcoded (GFP+) cells and PCR amplify barcodes 6. Sequence BCs by NGS to reveal the population architecture 36 4.1.1 Multiplicity of infection of the LV-BC. To use the cellular barcoding technology under controlled conditions, we first determined the multiplicity of infection (MOI) of each LV-BC batch produced. The MOI generally represents the ratio of infection agents (e.g. phages, virus or bacteria) to target cells. For example, when referring to a group of cells inoculated with virus particles, the MOI is the ratio between the number of virus particles to the number of target cells present in a defined space. MOI is determined based on the assumption that the infection agents associate with target cells in a fasion estimated by the Poisson distribution. Poisson distribution is a discrete probability distribution, specifically the probability of a given number of events occurring in a fixed interval of time or space if the events occur with a known average rate that is independent of the time since the last event. The formula to determine MOI is: (# of cells plated) x (GFP%/100) x dilution factor (volume per well/volume of virus) The MOI is further known for some standard cell lines, which are routinely used to determine the MOI of lentivirus batches. We used HEK293 cells that for a transduction of ~100% needs a MOI of 5. In order to calculate the MOI of our batch, we needed to titrate our lentivirus in a dilution series to obtain a transduction efficacy in a range between 80-30% to then calculate the average of all obtained MOI’s in these conditions. In our experiment, the average of all MOI for transductions in the range of 80-30% was 3.1 (Table 2). From previously experiments, we knew that this MOI would be sufficient to use for transductions of the hAML cell line HEL, since aiming for <10% efficiency in order to avoid multiple LV-BC integrations in individual cells. Table 2 – Percentage of transduced cells for each LV-BC titration and respective MOI. Dilution factor 1:8 1:16 Td Cells (%) 69.6 39.4 MOI Average MOI 2.8 3.2 3.1 37 4.1.2 Optimization of the transduction protocol to define conditions that fulfill the quality controls required to implement the cellular barcoding technology. As for many other advanced technologies, implementation of cellular barcoding was closely linked to diverse technical and analytical challenges. Here, I will pinpoint those challenges and explain how we controlled them. Two types of events can occur that impair our ability to interpret the data conclusively: multiple integration (during transduction) and repeat use of barcodes (overloading during the initiation of the selection system). Multiple lentiviral integrations into one cell distort barcode analyses by falsely indicating that the descendants of multiple cells in the starting populations share the same fate, when it is in fact rather only the progeny of one cell, which is labeled with two (or more) different barcodes, that is traced. To minimize the probability of multiple integrations and thus the ratio of cells carrying multiple, relative to a single, barcode(s), one must target a low transduction efficiency, generally <10%. The transduction conditions thus have to be optimized suitably for every system as it depends on the amenability of each cell type to LVBC transduction. If one does not realize and correct for that certain cells carry multiple BCs, the conclusions drawn may be seriously flawed. However, even when multiple integrations do occur within one cell, this is a relatively minor problem if noticed and managed bioinformatically. Any two (or more) barcodes that always covariate very closely in sequential samples, most notably if these include several different organs, must simply be bundled (i.e. we just consider one of them and disregard all NGS-reads of the shared BCs). The adverse impact on the overall analyses will thereby be limited to that a fraction of the traceable number of BCs in the library would be lost, leading to lower maximal resolution of the system but no errors in the conclusions inferred. A second type of problem arises if the same barcode is represented by independent cells. Such repeated usage of barcodes is due to overloading of the system and forms a more significant concern, as it would result in false assignments of lineage relationships. We estimated the probability/expected frequency of repeat use of barcodes within one mouse by transferring LV-BC transduced cells from the same batch into at least two separate mice, followed by subsequent comparison of the barcodes recovered from those mice. If the same barcodes are shared at higher levels than expected by chance in independent mice, the risk is also significant that there is repeat use of a barcode in independent cells within a mouse. The higher degree of correlation in barcode usage between two mice, as measured by e.g. the Pearson correlation coefficient, the greater the risk of repeat use within each of them. As 38 a general guideline to help limit repeat use, the number of cells tracked per mouse should represent max ~10% of the total library diversity. Repeat BC usage depends on the frequency of LV-BC with the same barcode within the library, the transduction efficiency, the size of the library, the number of labeled cells injected, and the engraftment probability. For the design of an experiment, the level of transduction and number of labeled cells transferred are the parameters we readily can alter and these should thus be optimized and/or assayed in titration so that the risk of both multiple integrations and repeat use is limited. A final step to increase the ability to do conclusive lineage tracing analyses using the cellular barcoding system is to minimize the degree of cell division between transduction and injection. When barcoded cells divide we will generate several cells with same barcode (not a real problem in it self, since they are related/of the same lineage), but if some of these divided cell-lineages preferentially establish, the same BC will be found in the two tested mice and we cannot distinguish this from two independent cells with the same BC establishing, and we thus also have a high risk of repeat BC use in the same mouse. We therefore sorted for GFP+ and injected the transduced cells as soon as possible to avoid cell division. This means that a culture transduced at ~10% efficiency (which will be detectable by GFP+ only after several days) must be sorted when only ~2-3% of the cells can be identified as barcoded by the GFP+ tag, meaning that we can only trace 1-2% of transduced cells. In order to minimize the degree of multiple LV-BC integrations, we needed to understand the transduction efficiency of our system. We did a transduction with two serially diluted LVs that each encoded a different fluorescent protein (GFP and RFP, respectively), to define the transduction level (below ~10%) that yielded relatively few double positives (DP), i.e. cells transduced by two different viral particles, relative to the total transduction level (Figure 6). The blue line initially drops with the red, meaning that as total transduction (DP+GFP+RFP) decrease, and so does the relative frequency of DP cells. However, the blue line goes flat after the 1:4 dilution, meaning that decreasing the total transduction beyond ~5% do not yield any further decrease in the relative fraction of DPs (Figure 6). Based on these results, we subsequently aimed to always have transduction levels <5% (as measured by the peak of GFP+), in order to avoid multiple integration. As we also knew the importance of avoiding cell division between LV-BC transduction and injection (detailed above), we sorted at ~22h after transduction and can thus only expected to inject ≤1%. This clearly mean we would need to transduce very big numbers of cells, in order to obtain enough barcoded cells for in vivo experiments. This technical issue 39 led us to next optimize our transduction protocols to enable transduction of as much cells as possible in the available lab conditions. 1:1 1:2 1:4 1:8 1:16 0 5 10 15 20 25 30 LV -B C dilution factor % of tran sdu ced cells Total Transduced (D P+G FP +R FP) % D P of total Td Figure 6 – Effect of LV-BC dilution on the percentage of double positive (DP) and totally (DP+GFP+RFP) transduced cells. The bars represent standard deviation of the mean of two experiments at each point. The optimization of the transduction protocol focused on three key parameters: − The amount of cells per well during the transduction – aiming to have as much cells that the limited contact surface between the LV-BC and the cells allowed during the centrifugation. − The amount of polybrene added – this cationic polymer is used to increase the efficiency of infection of certain cells with lentivirus in cell culture. Polybrene acts by neutralizing the charge repulsion between virions and sialic acid on the cell surface. However, polybrene is also toxic for the cells after some time of contact, requiring optimizations of the amounts used to achieve enhanced transduction efficiency in the absence of severe toxicity. − The amount of LV-BC used – the more cells added the more LV-BC is required to achieve the same amount of transduction. The right balance between these aspects give us the maximum feasibility to transduce a total of 1.2 billion cells, to after sorting for the ~1% of GFP+ cells get the required number of cells needed for the xenograft studies. 40 Figure 7 – Optimization of the transduction protocol. A) Effect of different amount of cells per well during transduction on the % of transduced cells by time (used 1:10 LV-BC dilution). B) Effect of polybrene concentration during on different cell amounts versus the efficacy of transduction (used 1:10 LV-BC dilution). C) Effect of the amount of LV-BC at the time of transduction on the % of transduced cells over time. D) The % of transduced cells over time in the experiment identical to that used to generate the cells used in vivo, i.e. using the final conditions selected after transduction optimization (used 1:12 LV-BC dilution). We observed that the more cells we added to each well, the lower the frequency of transduction obtained. As we needed a low transduction and as high as possible numbers of transduced cells, we selected the maximum cell numbers that we had tested per well, between 3-6 million (Figure 7A), for further optimizations. To decide the respective concentrations required to promote an optimal interplay between the specific cell number used and the amount of polybrene, we titrated both parameters against each other (Figure 7B). We found that a doubling of the polybrene concentration to 12g/ml dramatically improved transduction efficacy, in both tested cell concentrations. However, a further doubling of polybrene to 24g/ml did not further improve the transduction to the same extent, and when using 3 million cells showed nearly no effect at all (Figure 7B). Based on these results, we decided to use 6 million cells per well with 0 50 100 150 0 1 2 3 4 5 6 7 8 Tim e (h) % of transduced cells 1:12 1:10 1:8 0 20 40 60 80 100 0 5 10 15 20 Tim e (h) % o f transduced cells 0.75M 1.5M 3M 6M 0 45 90 135 180 225 270 0 1 2 3 4 Tim e (h) % of tran sdu ced cells 0 6 12 18 24 30 0 2 4 6 8 10 P olybren e concen tration - g/m l % of tran sduced cells 3M 6M A B D C 41 12µg/ml of polybrene, as it gave maximal cell numbers together with ~5% transduction efficacy. In order to avoid multiple LV-BC integration, we aimed for transduction of strictly ≤5% of the total cells. The last condition to be optimized was thus the amount of virus we needed to obtain the desired percentage of transduced cells. These optimization experiments showed that a LV-BC dilution of 1:12 resulted in ~3% of transduced cells, while 1:10 resulted in ~4% and 1:8 in ~6,5% transduction (Figure 7C). Based on these results we decided to use a LVBC dilution of 1:12 as our chosen final condition, to guarantee a transduction level of less than 5%. To finalize, we performed an experiment incorporating all the optimized conditions, where at 135 hours after the transduction of 6 million cells per well, with 12µg/ml of polybrene and a LV-BC dilution of 1:12, we got a peak of 3,5% transduced cell, meaning that our optimization was successfully concluded (Figure 7D) 4.2 New method to determine L-IC frequencies in AML. Once all aspects of the transduction procedures had been optimized, we proceeded with the in vivo experiments. Importantly, our lab had already established the novel xenotransplantation mouse model of hAML cells that enabled us to pursue the proposed studies. Thus, we IBM xenotransplanted the FACS-enriched barcoded HEL cells in immunodeficient Rag2-/-c-/- mice. In detail, 6-8 week old mice were sub-lethaly irradiated (250 cGy, and subsequently kept on oral antibiotics, Bactrim) and 24h later injected with a 10x dilution series (down from 100,000 to 100) HEL cells, directly into the right tibia bone marrow (IBM injection) and then monitored daily for signs of disease. The whole process is schematically described in figure 8. Weekly quantification of tumor burden was done by flow cytometry-based evaluation of the absolute amounts of circulating GFP+ hAML cells in the PB. Animals were sacrificed when sick or 12-14 weeks after tumor injection and PB and Right Tibia (RT) (injected bone) as well as BM (other bones) were characterized by FACS. 48 cells, versus ~65% in the plate seeded with 500 cells, ~60% in 1000 plated cells and ~55% in 2000 plated cells (Figure 14C). Figure 14 – Tumor BC numbers and heterogeneity in vitro. A) Pie chart representation of the barcode architecture of 4 dose dependent set of duplicate biological samples, obtained in vitro after seeding LV-BC Td HEL cells. Barcodes are listed and color coded (red – most frequent BC, etc) for each individual pie chart. B) Number of barcodes detected on average in the seeded LV-BC Td HEL cells. C) The relative dominance of the top 10% most frequent BC as a fraction of the total number of barcoded cells in each well. To characterize the effect of the selective pressure of the in vitro and in vivo systems, we performed a direct comparison between the in vitro and in vivo system seeded/injected with the same number of the same batch of cells (1000 cells). This direct comparison allows us to characterize the relative effect that the selective pressure of the in vitro and in vivo systems have on an initial number of 1000 LV-BC Td HEL cells. Mean: 31 BCs Mean: 139 BCs Mean: 236 BCs Mean: 397 BCs B C 500 cells in vitro seeded 1000 cells in vitro seeded 125 cells in vitro seeded 2000 cells in vitro seeded A 49 Figure 15 – Effect of microenvironment pressure on 1000 LV-BC Td HEL seeded/ injected cells. A) Quantification of total number of BCs established in vitro/in vivo. B) Quantification of the contribution for total tumor by the top 1% most frequent BCs in each in vitro/in vivo sample. C) Quantification of the number of seeded/injected cells necessary to establish 1 BC in vitro/in vivo. We observed that from an initial number of 1000 LV-BC Td HEL a mean of 236 BCs established in vitro while only a mean of 17 BCs established in vivo (Figure 15A). We also observed that the barcode architecture established in vivo shows a significant skewing compared to the in vitro condition as confirmed by the fact that the top 1% most frequent BCs account for an average of ~70% of the total tumor in vivo while this number drops to ~15% in vitro (Figure 15B). To confirm the high selective pressure exerted in vivo we quantified the average number of cells necessary to be seeded versus injected to establish 1 single BC either in vitro or in vivo. We observed that we on average needed ~4 cells to establish 1 BC in vitro while in vivo this number rises to 231 cells (Figure 15C). To characterize the selective pressures mediating the limited in vivo establishment of LV-BC Td HEL cells we next compared the correlation between the number of cells seeded/injected in vitro/in vivo and the respective BC number established. We observed that the in vitro cultures fitted strongly to a linear correlation between the number of seeded cells and the number of BCs established (R square of linear fit = 0.9834) (Figure 16A). However, in vivo the relation between the number of injected cells and the number of BCs established was exponential (R square of semi-log fit = 0.9176) rather than linear (R square of linear fit = 0.7665) (Figure 16B), thus indicating that in contrast to in vitro, in the in vivo system we need an exponential increase in the cell number injected to keep a linear increase in BC numbers. A B C 50 Figure 16 – Effect of the selective pressure of the in vitro and in vivo systems. Quantification of the number of established BCs per number of cells seeded/injected in vitro (A) and in vivo (B). Dashed line depicts the fitting of the data to a linear distribution and solid line depicts the fitting of the data to a semi-log distribution. Linear if R square value > 0.95. 125 500 1000 2000 0 100 200 300 400 500 N um ber of plated cells N um ber of B C detected 1/5 1/4 1/3 1/4 Figure 17 – Quantification of L-IC frequency. A) Number of barcodes detected for each amount of cells plated in parallel in vitro cultures. B) Number of barcodes detected and L-IC frequency for each amount of cells injected into mice. We found that in vivo, 1,000 injected cells into mice compared with 10,000 cells, the higher number, dramatically increased the number of barcodes detected after disease establishment, however a further increase of 10x cells did not further improve the number of barcodes detected (Figure 17B). Looking to L-IC frequency versus number of injected cells, we were able to calculate the exact number of initial cell number needed to inject in order to obtain one L-IC established. These calculation showed that 1/59 L-IC established in an injection of 1,000 transduced HEL, while 1/99 in 10,000 and 1/518 in 100,000 (Figure 17B). In contrast, from in vitro data, we observed a proportional relation between the amounts of A B A B 51 plated cells versus the number of detectable barcodes: the more cells we plated, the more barcodes we had (Figure 17A). In relation to BCs establishment, in all different amount of plated cells (125, 500, 1000 and 2000) we saw that for each ~4 cells plated, 1 BC established (Figure 17A). L-IC establishment or frequency can also be measured by ELDA knowing how many cells were injected into each group; and that one of our groups failed to obtain establishment. We then also crossed these numbers with the ones given by NGS-based quantitative analysis of the frequencies of the different barcodes and compare their relative sensitivity. Table 3 – Results of ELDA, the software application for LDA. L-IC frequency was calculated for each amount of cells injected into mice. Confidence interval of [1493;116] with an estimated value of 417 cells to establish one L-IC. Cell # injected # of tested animals # of sick animals L-IC Frequency Estimative 100.000 2 2 1:417* 10.000 4 4 1000 3 3 100 3 0 * likelihood ratio test for single-hit model – p value of 0.156 > 0.05. From LDA results, we obtained disease establishment in all mice injected with cell numbers between 1,000 and 100,000 (Table 3). Considering L-IC frequency versus number of injected cells, the ELDA software gave us an estimate L-IC frequency of 1:417, meaning that in order to obtain one L-IC established, we needed to inject 417 cells (Table 3). These results revealed an increased sensitivity, as well as the dependence on the injected dose (potentially reflecting an influence of the engrafted niche), of cellular barcoding when compared to ELDA for the estimation of L-IC frequencies in AML xenograft models. 52 4.3 Screening for an AML stem cell-like phenotype for xenograft studies We were able to successfully create an in vivo xenograft model where we could study L-IC dynamics. However, L-IC are also viewed as the population responsible for the maintenance of the disease, metastases and relapse and as we want to optimize the potential translation of future therapeutic observations in similar xenograft studies, we need to search for an hAML cell line that better than HEL represents a clinically relevant phenotype in terms of expression of stem cell markers. For that reason, we screened 3 different AML cell lines, using CD34 and CD38 as stemness markers, in order to have a L-IC population characterized as CD34+CD38-, figure 18. Figure 18 – hAML screening of CD34+CD38population in 3 hAML cell lines, GDM-1, HEL and KG-1. The cell line with the most similar CD34+CD38profile to primary human patient samples was the GDM-1 cell line with a big percentage of non-L-IC population (~99%) and with a small fraction of L-IC (<1%). HEL cell line also had a big amount of non-L-IC (~99%), however, it virtually lacked cells stainig positive for CD34 (Figure 18). KG-1 showed the higher amount of L-IC population (~72%), between the three cell lines, being the one with less similarities with primary human patient profile (Figure 18). In order to have a way of specifically target L-IC of the cell line, we needed to have some molecular marker known as to be more expressed in the fraction of CD34+CD38cells. CD47 is expressed on multiple human normal cell types and tumors and it’s known that the SIRPα engagement by CD47 critically inhibit phagocytosis. Blocking CD47/SIRPα interactions, turns off the “don’t eat me” signal and favors phagocytosis, with dramatic effects on leukemia CD34 CD38 CD34 CD38 CD38 CD34 HEL GDM-1 KG-1 53 xenograft models (Majeti et al. 2009; Majeti 2011). It is known that CD47 expression is an adverse clinical prognostic factor and expressed in the tumorigenic CD34+CD38fraction. Accordingly, CD47 targeting have shown promising results in xenograft models of hAML (Majeti 2011). It was also used as therapeutic antibody to target human AML stem cells clinically. For these reasons, we screened CD47 expression in the cell lines with most similar “stemness” profile to phAML samples. C D 34-C D 38+ C D 34+C D 38C D 34-C D 380 20 40 60 80 100 120 H E L M FI betw een S tem n ess vs % C D47+ G eoM ean M ean C D 34-C D 38+ C D 34+ C D 38+ C D 34+ C D 38C D 34-C D 380 10 20 30 40 G D M -1 M FI betw een S tem ness vs % C D 47+ G eoM ean M ean Figure 19 - Expression of CD47 in AML cell lines. A) Expression of CD47 and its isotype and the ratio of MFI between CD47/Isotype staining in both cell lines. B) Ratio of MFI between CD47/Isotype staining in each population of CD34 and CD38 in both cell lines. When we compared the overall expression of CD47 between GDM-1 and HEL, the one who expressed more is HEL cell line (Figure 19A). Then we analyzed its expression in individually population of CD34 and CD38 cells, between both cell lines, and it had some variations. In the subpopulation of L-IC (CD34+CD38-), the expression of CD47 was relatively higher in GDM-1 than in HEL (Figure 19B). In fact, among all the subpopulations of GDM-1 evaluated the CD47 expression was highest in the CD34+CD38fraction (Figure 19B). H E L G D M -1 0 50 100 150 R a tio o f M F I b etw e en C D 4 7/Is o ty p e s ta in in g G eo M ean M ean HEL GDM-1 CD47 Isotype Unstain A B 54 As the cell line with a higher expression of CD47 in the L-IC population fraction was GDM-1, we choose to continue the model with this line. GDM-1 had the most similar “stemness” profile to primary human patient samples, among all thecell lines screened. However, as we never used this cell line before, its capacity of to be transduced and its behavior in vivo, was unknown. Therefore, we used it in a small experiment just to study these aspects. We started by transducing the cell line with our LV-BC with an MOI of 3,7. We used 500,000 cells per well and a LV-BC dilution of 1:18. As control, we used HEL cell line, also with 500,000 cells per well with a LV-BC dilution of 1:10. 15H 33H 4days 11days 0 20 40 60 80 100 G DM -1 Td vs H el Td H EL LV -B C D ilu tion 1:1 0 G D M -1 LV -B C D ilu tio n 1:1 8 Figure 20– Kinetics of the transduction of GDM-1 and HEL as control. Both cells line were successfully transduced, HEL with a transduction peak of ~80% of the total HEL cells and GDM-1 with a peak of ~50%. However, the transduction kinetics were different between them. Between 15-33 hours, HEL, as expected, had a continous increase of transduced cells (~15% to 50%). In contrast, GDM-1 showed almost no transduced cells. Between 33h and 4 days, HEL continued to have an increase of transduced cells (~50% to 80%), and GDM-1 increased from ~0% to 50% of the cells (Figure 20). In order to just understand, at an initial phase, how GDM-1 behaved in vivo, we did not worry about keeping the percentage of transduced cells low (to avoid multiple integrations). We used the transduced GDM-1 at transduction peak of ~50% and injected these into mice (using the same conditions as previously described for HEL). Once the disease established we followed its progression, by measuring the number of GFP+ cells in PB weekly. Animals were sacrificed when sick or when we detected >50000 cells/ml in the 55 PB after tumor injection and PB and RT (injected bone) as well as BM (other bones) were characterized by FACS. Figure 21 - LV-BC transduced GDM-1 versus LV-BC transduced HEL injected mice. A) Number of hAML cells per ml of PB in each LV-BC transduced GDM-1 and LV-BC transduced HEL injected mouse. B) Total number of hAML cells per organ in each LV-BC transduced GDM-1 and LV-BC transduced HEL injected mouse. We observed that GDM-1 transduced cells showed slower leukemic progression compared to HEL transduced cells. All GDM-1 groups were engrafted at week five and HEL groups at week four, however, the lowest group of GDM-1 was 2000 cells injected, while HEL group was 100 cells (Figure 21A). In relation to disease establishment, in GDM-1 just three mice groups established the disease (10,000, 50,000 and 125,000 cells), the others died before appropriate analyses could be done. In contrast, the four higher groups of HEL obtained disease establishment (Figure 21A). We also found that GDM-1 LV-BC Td showed lower overall tumor burden at disease threshold (80000 cells/ml) compared to HEL LV-BC Td (Figure 21B). GDM-1 also showed an incidence of leg tumor (LegT before high PB levels) of: 100% in 2000 group (1 animal), 33% in 10000 group (3 animals), 50% in 50000 group (2 # # # # N = 4 in all groups PB quantification of BC Td HEL Organ quantification of BC Td HEL B A 56 animals), 100% in 250000 group (3 animals) and 0% in 1250000 group (3 animals) with an overall rate of 42%. Since the GDM-1 cell line, despite its interesting phenotype, proved to be less adequate to use in our model than the HEL cell line; and given that our ultimate goal is to use (clinically relevant) phAML, we decided to proceed to the application of our in vivo model and barcoding technology with AML patient blasts. 4.4 Proof of concept with primary AML patient samples. Having shown the feasibility of establishing in vivo a LV-BC Td hAML cell line and confirmed its potential to address relevant aspects of tumor biology, we next sought to apply our optimized in vivo procedures to establish a preclinical system to study L-IC dynamics of clinically relevant AML cells, i.e., we tried to use primary human AML (phAML) patient blasts as the recipient cells for our barcodes. Specifically, BM samples collected at diagnosis, from patients admitted to IPO-Lisboa, were used to isolate the mononuclear cell fraction by ficoll gradient before freezing. Upon thawing, we depleted CD3+ cells, before transducing phAML with the LV-BC. Barcoded cultures were then characterized (hCD33, hCD13, hCD45, mCD45, live/dead) over time, in vivo and in parallel in vitro cultures. For the in vivo studies we first tried the Rag2-/-γchain-/- xenograft model, which had successfully accommodated barcoded variants of the HEL hAML cell line. Although immunodeficient in many aspects, Rag2-/-γchain-/- mice were not sufficiently permissive to engraftment of phAML cells for our studies. We thus moved on to try the more xenoengraftment-permissive NSG mice. These started to take oral antibiotic (bactrim) one day before being sublethally irradiated (250cGy), which in turn was done the day before tumor inoculation. The injection was IBM in the right tibia, with mice under anesthesia. Once the disease established we followed its progression, by measuring the number of GFP+ cells in the PB every two weeks. Animals were sacrificed when sick or 12-14 weeks after tumor injection, and PB and RT (injected bone) as well as BM (other bones) were characterized by FACS. Cells were lysed and genomically integrated barcodes amplified by a nested PCR and then sequenced by NGS, to reveal the barcode composition of each population. We were thereby able to trace how many single-cancer-cell-derived lineages established in each mouse and thus how many L-ICs contributed to establishing the disease in each animal. This whole process is schematically described in figure 22. 57 4.4.1 Transduction of phAML with LV-BC. As discussed for the barcoding of HEL cells, the successful clearance of the quality controls of the barcoding system is dependent on the LV-BC transduction conditions. While all patient samples are largely unique. Thus makes it much harder to transduce than the HEL cell line, so lentivirus with high MOI was used. Figure 22 – Scheme of the preclinical system to study L-IC dynamics from primary human AML samples, with the experimental process and chronologically description of each task. Tasks are from 1 to 6. As these patient samples are heterogeneous, we also needed to have a different strategy to gate the cells when analyzing the transduction levels, to guarantee that we were looking at leukemic cells. We started to select human cells from mouse cells using anti-CD45 for each species as a two markers. After excluding cells staining for the live/dead marker, we used anti-CD13 and/or CD33 to select myeloid cells for further analyses (Figure 23A), where we then analyzed GFP positivity, stemming from co expression with our barcode sequence. Specifically, we were able distinguish between transduced and untransduced cells using 488nm laser and the standard channel to detect GFP expression (Figure 23B). We successfully transduced all phAML samples and transduced cells consistently reached 1. PB / BM sample: isolation of MNC fraction and CD3 depletion - primary human (ph) AML 0102103104105 0 102 103 104 105 Mouse sceening panel phAML21 5 colorsЙ<APC-A>, <APC-Cy7-A> subset <PE-Cy7-A> <PE-A> AML patient at diagnosis 2. Transduction of phAML samples and FACSorting with its characterization (hCD33, hCD13,hCD45, mCD45, live/dead) 3. IBM xenotransplantaion into imunodeficient mice CD33 CD13 NSG (NOD-SCID gamma chain KO) mice, irradiated (200 rads), w/ Antibiotics 4. FACS PB monotorization of phAML cells every 2 weeks CD33 CD13 5. Sacrifice animals (sick or 1214 weeks after IBM) and FACS characterization of phAML samples in mouse PB and RT (injected bone) and BM (other leg bones) and PCR amplify barcodes 6. Sequence BCs to reveal the population architecture 64 65 5. Discussion Intratumor heterogeneity is a major driver of the disease recurrence often observed after cancer therapies. The frequent relapse of AML after standard chemotherapy is a prototypical example of how the emergence of treatment-resistant tumor escape variants efficiently limits clinical benefit of available anti-cancer drugs (Hughes et al. 2014). While the dominant mutations that contribute to AML heterogeneity and relapse have been characterized, little remains known about the mechanisms that govern sub-clonal dynamics/evolution and their impact on the response to treatment and recurrence. L-ICs are considered responsible for the maintenance and relapse of AML. This rare subpopulation of leukemic cells predominantly display a phenotype positive for CD34 and negative for CD38 (Lapidot et al. 1994). We have combined optimized intra-bone marrow injection-based xenotransplantation models with a cutting-edge technology for lineage tracing of hAML populations, to establish a preclinical experimental system allowing us to study the frequency and clonal dynamics of L-ICs at unprecedented depth. Our lab has recently established an in vivo xenotransplantation model in which hAML cells are IBM-injected in the right tibia of sub-lethally irradiated immunocompromised mice. This model consistently and accurately reproduced the leukemia growth, spread and associated mortality observed in AML patients (Figure 9) and was thus employed for our studies. A key aspect of this study was to adapt the cellular barcoding technology (employed in collaboration with the group of Ton Schumacher), to the context of tumor development. Specifically, we aimed to enable quantification and analyses of the dynamics of multiple parallel L-ICs in our in vivo xenotransplantation model of hAML. Importantly, we have been able, for the first time, to trace hundreds of phAML populations interacting in vivo, each stemming from an individual primary human AML cell. As the cellular barcoding technology was originally developed to lineage trace murine immune cells (Schepers et al. 2008; Naik et al. 2013), we had to optimize the existing protocols for lentiviral-based barcode transduction to meet the specific conditions required to use this approach to trace the dynamics of single-hAML-cell lineages in our xenograft model. With the optimized protocol (Figure 7), we successfully achieved conditions compatible with passing the required quality controls and thus proceeded with the in vivo experiments. After the NGS-based DNA sequencing of each sample, quality reads were processed in a bioinformatics pipeline (Naik et al. 2013) that yielded a quantitative map of the relative frequencies of each and every barcode, with respect to the total number of barcode reads, 66 detected in the BM of pairs of mice injected with 1,000, 10,000 and 100,000 LV-BC Td HEL cells. After we ensured that the sampling and distribution controls were met, we observed a total average number of 17, 101 and 193 barcodes, respectively (Figures 12 and 13). This indicated that the amount of barcodes that established in vivo positively correlated with the amount of LV-BC Td cells injected (Figure 13). Importantly, the number of established barcodes in all tested groups was lower than 10% of the total library (i.e. 260 barcodes), thus supporting the suitability of applying the BC technology in our in vivo model (Figure 12). Overall, the data discussed above showed that we could transduce hAML cells with barcode lentiviral vectors, establish the resulting GFP+ cells in vivo and pass all the quality control steps, confirming the technical feasibility of using the BC technology, in a way that can meet the required quality controls, to assess single-tumor-cell-lineage dynamics in our hAML xenotransplantation model. We next sought to validate the use of this system to address relevant biological questions in cancer biology. To this end, we designed experiments that allowed us to measure the relative extent of selective pressure exerted on the LV-BC transduced hAML HEL cell populations upon xenotransplantation versus continued in vitro growth. As a direct comparison of the degree of selective pressure exerted by growing LV-BC transduced HEL cells in vitro (on plastic in 21% oxygen) versus in vivo (as xenografts in Rag2-/-γc-/- mice) we evaluated the BC architecture after establishment of an equivalent batch of cells in each system. We showed that in vivo establishment of 1,000 hAML cells implied a significant reduction of the number of barcodes found and the relative representation of their diversity, due to in vivo skewing of the barcode architectures, as compared to parallel in vitro culture establishment (Figures 15 and 16). As expected, this data supported that the degree of selective pressure exerted by growing LV-BC Td HEL cell line cells in our in vivo model was higher compared to that during in vitro growth. Besides confirming the technical feasibility of using the BC system in our model, these results confirmed that we could directly assess the sub-clonal dynamics upon variable selective pressures and thus showed the potential answer biologically relevant questions related to these aspects. Notably, the amount of injected cells itself appeared to affect the relative number of barcodes generated per initial cell. Specifically, we found that the initial dramatic increase in barcodes detected after in vivo disease establishment, between mice injected with 1,000 versus 10,000 cells, did not continue to the same extent when a further equivalent 10x increase of the number of injected cells (to 100,000) was evaluated (Figures 13, 17). By contrast, the relative increase of the number of barcodes establishing in parallel in vitro 67 cultures was similarly proportional for the entire analyzable range of the system (Figures 14, 17). These data suggested that the number of hAML cells injected/ assayed in each mouse itself was a key determinant for the L-IC frequency (Figure 17), which must be considered when designing such experiments. We calculated the average number of injected cells needed to obtain one barcode established, as a measure of the L-IC frequency, for each cell concentration tested (Figure 17). L-IC frequency can alternatively be measured by ELDA, based on the fraction of mice that establish disease in groups injected with decreasing amounts of cells (Table 3) (Hu & Smyth 2009). When we cross-compared the L-IC frequencies obtained from the ELDA with those given by our NGS-based quantitative analysis (of multiple parallel barcodes, for all the different groups injected with different cell numbers), the later showed superior sensitivity. Specifically, in the mice injected with 1.000 LV-BC transduced HEL cells, the 17 BCs detected on average correspond to a L-IC frequency of 1/59 cells (Figure 17), much more frequent than than estimated by the ELDA software that gave us a predicted L-IC frequency of 1/417 cells (Table 3). Considering that barcoding-based L-IC frequencies dropped when mice were injected with more cells (Figure 17), we speculate that the ELDA consistently underestimates L-IC frequencies. The underlying reason for the too low predictions is the inherent need of ELDA to consider multiple cell concentrations in the regression analyses (Hu & Smyth 2009), including some conditions with higher numbers of hAML cells than the lowest yielding engraftment in a fraction of the animals, where the L-IC frequency is relatively highest according to our barcoding-based L-IC analyses (Figure 17). Our results thus indicated that the use of cellular barcoding for quantifying L-IC is potentially both more sensitive and accurate than ELDA, due to its ability to measure multiple parallel L-ICs and only consider the most optimal cell number tested. Of note, the barcoding-based L-IC frequency only got down to the range of the ELDA estimate (1/417) when mice were injected with 100,000 cells (1/518), which corresponded to plateauing barcode numbers due to niche saturation (Figure 16). Since L-ICs carry clinical value as predictive markers of relapse (Kreso et al. 2013; Kreso & Dick 2014) we next aimed to optimize the potential translation of future therapeutic observations made in similar xenograft studies. To this end, we focused our efforts on searching for a hAML cell line that better than HEL represented a clinically relevant phenotype in terms of expression of stem cell markers. We screened 3 hAML cell lines for a phenotypic “stemness” profile (CD34+CD38-) most similar to that of phAML samples and settled on GDM-1 (Figure 18). From the literature, we knew that CD47 expression was an 68 adverse prognostic factor on patient’s phAML blasts, (Majeti et al. 2009). It has further been suggested as therapeutic antibody target on human AML cells and have demonstrated efficacy against AML L-IC in xenotransplantation models. (Majeti 2011). Therefore, we next screened the HEL and GDM-1 cell lines for their CD47 expression and found that the later expressed the relatively highest levels of CD47 on the CD34+CD38subpopulation and thus represented an interesting model for preferential targeting of this disease-driving subpopulation (Figure 19). We therefore pursued with xenotransplantation of LV-BC transduced GDM-1 cells. We observed that LV-BC transduced GDM-1 cells showed slower leukemic progression compared to the corresponding HEL cells and also found that GDM-1 LV-BC transduced cells generated lower overall tumor burden compared to LV-BC Td HEL cells (Figure 21). Another major drawback was that 42% of the mice injected with GDM-1 developed sizable tumors in the leg muscle, requiring animal sacrifice, before developing systemic disease. The continued use of the GDM-1 cell line would thus require us to have much more mice per group than planned in order to compensate for such loss. From an economical and ethical cost/benefit point of view, the GDM-1 cell line has thus shown to be far inferior to the HEL cell line for the use in our model. Rather than relying on alternative cell lines to HEL to provide a clinically relevant phenotype (in terms of expression of stem cell markers), we instead focused our efforts on directly implementing phAML in a further optimized xenograft model. Following a series of optimizations of both the phAML grafts (e.g. T cell depletion) and the recipient mice (e.g. the use of NSG rather than Rag2-/-γc-/- mice), our lab has defined conditions that allowed robust establishment of xenografted phAMLs. We used this model and for the first time established in vivo barcoded phAML samples. We showed that phAML cells could be readily LV-BC transduced, without loosing their potential to engraft or expand in vivo (Figure 24). In detail, we LV-BC transduced a total of 6 phAML and established these in vivo, which lead to phAML growth, spread and associated host mortality (Figures 23 and 24). The fact that the therapeutic responses (or lack thereof) in vivo of each phAML sample mirrored the clinical results in the corresponding patients supported that the barcoded phAML cells that established in our model retained physiologically relevant aspects of the original patient samples. Genome-integrated barcodes could further be readily recovered (amplified by optimized PCR-based protocols and deep sequenced) from GFP+ phAML cells isolated from the leukemic mice (Figure 25). 69 In contrast to the observations done when we used HEL, injection with more phAML cells yielded an even higher relative increase in the number of barcodes detected. Specifically, we found 5-fold more barcodes in the mouse injected with 6M cells, than the mice injected with 3M cells (Figure 25). It is thus possible that while sub-clones appeared to compete with each other (leading to reduced L-IC frequencies) when we increased the number of HEL cells transplanted into mice, they instead synergized with each other (leading to increased L-IC frequencies) when we increased the number of injected phAML cells. In both systems, the amount of cells injected in each mouse thus itself impacted the L-IC frequency observed, but in different directions (for the ranges tested). When we measured the relative dominance of the top 1% most frequent BC, of the total barcoded tumor cells, we observed that >2-fold more of the tumor was represented by this fraction of the single-celllineages in the BM in the mice injected with 3M LV-BC transduced phAML 22 cells compared to the mouse injected twice as many similar cells (Figure 25). This indicated that the amount of phAML cells injected in each mouse itself is a key determinant also for the BC family dominance/ representation of the tumor heterogeneity, in a similar way as previously observed between the 1,000 cell and 10,000 cell groups in the HEL-based hAML model. In conclusion, we have established models that allowed studies of the L-IC dynamics of hAML cell lines as well as AML patient samples. The phAML model further provides other interesting applications, such as being used as a relevant therapeutic model for AML. Primary human AML samples naturally showed a normal “stemness” profile and we can further use anti-CD47 therapies to study its impact on the maintenance of disease. We can also study the impact of other new clinical drug candidates on the sub-clonal heterogeneity of the disease. The next layer of resolution will be the incorporation of the (epi)genetic and transcriptomic profiles, with the barcode architecture, in the tumors relapsing after treatment. Once well established, our model will have multiple applications in cancer research and, when combined with other cutting edges technologies such as single-cell-sequencing, it carries great promise and enormous power to enhance the understanding of how tumor cell population dynamics underpin cancer genetics. 70 71 6. References 72 73 6. References Alexandrov, L.B. et al., 2013. Signatures of mutational processes in human cancer. Nature, 500(7463), pp.415–21. Anderson, K. et al., 2011. Genetic variegation of clonal architecture and propagating cells in leukaemia. Nature, 469(7330), pp.356–361. Ben-Bassat, H. et al., 1982. Establishment and characterization of a new permanent cell line (GDM-1) from a patient with myelomonoblastic leukemia. Leukemia Research, 6(6), pp.749–752. Bonnefoix, T. & Callanan, M., 2010. Accurate hematopoietic stem cell frequency estimates by fitting multicell Poisson models substituting to the single-hit Poisson model in limiting dilution transplantation assays. Blood, 116(14), pp.2472–2475. Burrell, R. a. & Swanton, C., 2014. The evolution of the unstable cancer genome. Current Opinion in Genetics and Development, 24(1), pp.61–67. Cairns, J., 1975. Mutation selection and the natural history of cancer. Nature, 255(5505), pp.197–200. Caldas, C., 2012. Cancer sequencing unravels clonal evolution. Nat Biotech, 30(5), pp.408– 410. Campbell, P.J. et al., 2010. The patterns and dynamics of genomic instability in metastatic pancreatic cancer. Nature, 467(7319), pp.1109–1113. Chan, W.-I. & Huntly, B.J.P., 2008. Leukemia stem cells in acute myeloid leukemia. Seminars in oncology, 35(4), pp.326–335. Cheng, L. et al., 2010. Mouse models for cancer stem cell research. Toxicologic pathology, 38(1), pp.62–71. Ding, L. et al., 2012. Clonal evolution in relapsed acute myeloid leukemia revealed by whole genome sequencing. Nature, 481(7382), pp.506–510. Doulatov, S. et al., 2012. Hematopoiesis: a human perspective. Cell stem cell, 10(2), pp.120– 136. Gerber, J.M. et al., 2012. A clinically relevant population of leukemic CD34+CD38cells in acute myeloid leukemia. Blood, 119(15), pp.3571–3577. Gerlinger, M. et al., 2014. Genomic architecture and evolution of clear cell renal cell carcinomas defined by multiregion sequencing. Nature genetics, 46(3), pp.225–233. Gerlinger, M. et al., 2012. Intratumor heterogeneity and branched evolution revealed by multiregion sequencing. The New England journal of medicine, 366(10), pp.883–892.