The ectopic expression of four pluripotency associated transcription factors [Oct4 (Pou5f1), Sox2, Klf4, and c-Myc] enables reprogramming of differentiated somatic cells into induced pluripotent stem cells (iPSCs) ( 1, 2). In comparison to other reprogramming means such as nuclear transfer or cell fusion, iPSC generation remains an inefficient process yielding not only fully but also partially reprogrammed cells with aberrant transcriptional profiles. Attempts to reduce the number of reprogramming factors (RFs) to three (3), two (4), or even one (5) demonstrate that only Oct4 is not dispensable. Because yield and speed are yet further decreased by such complete omission of factors, using all four RFs intending an equal stoichiometry remains the standard approach. A recent publication suggested by contrast a nonequimolar stoichiometry with a three-fold Oct4 excess for more efficient induction of pluripotency in human somatic cells (6) but did not investigate whether less optimal stoichiometries also result in less perfect reprogramming as measured by molecular analyses. However, we speculate that the estimation of actual transgene expression levels according to applied virus loads might be biased by differences in integration sites and copy numbers among various cells. In other words, it is not yet fully understood whether high amounts of Oct4-expressing viruses also lead to an evenly high Oct4 expression in each transduced cell. Various fluorescent proteins are available, which not only can be expressed in living cells but will also enable separate analysis by cytometry based quantification means, if appropriate filter combinations are applied. To this end, we chose a quantifiable gene transfer system consisting of bicistronic lentiviral vectors that coexpress the canonical RFs along with fluorescent reporters (7) such as eGFP, dTomato (a red fluorescent protein), Cerulean (a cyan fluorescent protein), and Venus (a yellow fluorescent protein). Thus, we were able to subdivide a randomly infected population of murine fibroblasts into populations with distinct RF expression profiles and could observe how these performed in generating putative iPSC colonies, which we further analyzed with regard to certain molecular and functional characteristics. Here, we demonstrate that a distinct RF stoichiometry, defined by high Oct4 levels opposite to modest ones for Sox2 and Klf4, is required for a satisfying iPSC generation efficiency and promotes the successful completion of the reprogramming process. RF stoichiometry influences both efficiency and quality of iPSCs and therefore has to be addressed in comparative studies, resulting in implications for novel reprogramming strategies such as viral delivery of the RFs with nonintegrating vectors or the direct employment of RF mRNAs or proteins.
Somatic cells can be reprogrammed toward pluripotency by overexpression of a set of transcription factors, yielding induced pluripotent stem cells (iPSCs) with features similar to embryonic stem cells. Little is known to date about stoichiometric requirements of the individual reprogramming factors (RFs) for efficient reprogramming and especially about whether stoichiometry also influences the quality of derived iPSCs. To address this important issue, we chose bicistronic lentiviral vectors coexpressing fluorescent reporters (eGFP, dTomato, Cerulean, or Venus) along with the canonical RFs to transduce a bulk of murine embryonic fibroblasts (MEFs). Using a flow cytometric approach, we were able to independently and proportionally quantify all fluorophores in multiple-infected MEFs and more importantly could sort these cells into all 16 stoichiometric combinations of high or moderate expression of the four factors. On average, we obtained about 600 alkaline phosphatase-expressing colonies from 20,000 seeded cells. Interestingly, only seven different stoichiometric ratios gave rise to any colonies at all. The by far most colonies were obtained from those fractions, where Oct4 was in excess over the other three factors (2,386 colonies/20,000 cells), or where both Oct4 and c-Myc were in excess over Sox2 and Klf4 (1,593 colonies/20,000 cells). Our findings suggest that increased Oct4 levels opposite to modest ones for Sox2 and Klf4 are required for satisfying reprogramming efficiencies and that these stoichiometries are also highly beneficial for achieving a stable pluripotent state independent of ectopic RF expression. Finally, the eligible Oct4high, Sox2low, and Klf4low subpopulation only resembles a small fraction of cells targeted by equal vector amounts, suggesting the necessity to address stoichiometry also in alternative approaches for iPSC generation or between different experimental systems. © 2011 International Society for Advancement of Cytometry.
MATERIALS AND METHODS
Cells and Culture Conditions
Medium for 293T cells (ATCC, Manassas, VA; Cat-#: CRL-11268) consisted of DMEM High Glucose supplemented with 10% fetal calf serum (FCS), 100 U mL−1 penicillin, 100 μg mL−1 streptomycin, 2 mM L-glutamine (PSG), and 0.1 mM β-mercaptoethanol (βME). Both murine and human fibroblasts were cultivated in DMEM Low Glucose with 10% FCS, PSG, and βME. MEF cells from mouse strain C3H were used for reprogramming experiments and also as feeder cells for mouse iPSCs after mitotical inactivation by 1.5-h treatment with 10 μg mL−1 mitomycin C (Sigma, St. Louis, MO). Irradiated (30 Gy) CF1 MEFs served as feeder cells for human iPSCs. Human fibroblasts had been derived from a fetal liver (cell line F-134, courtesy of Michael Ott, Hannover Medical School). Medium for cells to become or considered as mouse iPSCs contained Knockout DMEM, 15% FCS, PSG, βME, nonessential amino acids (NEAA, 0.1 mM each), and 1,000 U mL−1 LIF (in-house preparation). Human iPSC medium was composed of Knockout DMEM/Ham's F-12 (1:1) with 20% Knockout Serum Replacement (KOSR), PSG, βME, NEAA, and 4 mg mL−1 b-FGF (human recombinant, Peprotech, Rocky Hill, NJ). Unless otherwise noted, all cell culture reagents were purchased from Gibco (Carlsbad, CA) and PAA (Pasching, Austria).
For preparation of lentiviral vector supernatants, 293T cells were seeded at 50% confluency in 10-cm dishes 1 day prior to transient transfection by calcium phosphate precipitation in a HEPES-containing buffer. The RFs were cloned as described ( 8) based on a third-generation self-inactivating lentiviral backbone (pRRL.PPT.PGK.eGFPpre, kindly provided by L. Naldini, Milano, Italy) and equipped with NheI, AgeI, and SalI sites (9). We introduced the retroviral spleen focus-forming virus (SFFV) U3 promoter amplified as NheI and AgeI fragment. The human cDNAs encoding the RFs Oct4, Sox2, Klf4, and c-Myc were engineered with a Kozak sequence and amplified as AgeI and SalI fragments. To enable visualization of RF expression, we introduced fluorescent markers, namely eGFP (enhanced green fluorescent protein), Venus (YFP derivative), the red fluorescent protein dTomato, and the blue fluorescent protein Cerulean via an EMCV IRES (internal ribosomal entry site) sequence in the SalI site downstream of the RFs. The cells were exposed to 10 μg lentiviral vector plasmid, 10 μg pcDNA3.GP.4xCTE (expressing HIV-1 gag/pol), 5 μg pRSV/Rev, and 3 μg pMD.G (encoding VSV glycoprotein). Virus-containing supernatants were collected after 24 and 48 h, filtered (0.2 μm), and concentrated by centrifugation (6 h at 15,000 × g and 2°C). Titers were calculated by determining the fraction of fluorescing cells (by flow cytometry) in MEF cultures infected with dilution series of the concentrated virus suspensions 3 days after transduction.
Cells were dissociated with 0.25% Trypsin and 1 mM EDTA, resuspended in PBS (w/o Ca and Mg) with 1% FCS, and filtered (40 μm). Analyses of single and quadruple fluochrome transduced cells were performed with a LSRII cytometer (Becton Dickinson, Franklin Lakes, NJ) using the following instrument settings: for eGFP and Venus, we used a 488-nm laser for excitation and 510/21 or 550/30 bandpass filters for emission, respectively, separated by a 525LP dichroic mirror. A yellow laser (532 nm) was used for dTomato excitation (emission: 610/20) and a violet one (405 nm) for Cerulean (emission: 525/50). Detector voltages were as follows: eGFP 320 V, Venus 320 V, dTomato 371 V, and Cerulean 350 V. Automatic compensation setup was performed using cell lines expressing single fluorophores. The following compensation values were used: −70.1% Venus and −2.3% Cerulean for eGFP, −35.7% eGFP and −13.0% dTomato for Venus, −5.0% Venus for dTomato, and −2.1% eGFP for Cerulean. For cell sorting, we used a two-laser excitation on Aria IIu cytometer (Becton Dickinson) using 488-nm excitation for eGFP (510/20), Venus (550/30), and dTomato (610/20), separated by 525LP and 595LP dichroics, respectively, as well as a 405-nm excitation for Cerulean (emission: 550/40). Detector voltages were as follows: eGFP 500 V, Venus 500 V, dTomato 600 V, and Cerulean 550 V. The following compensation values were used: −44% Venus for eGFP, −88% eGFP and −40% dTomato for Venus, −5% eGFP for dTomato, and −5% eGFP for Cerulean. 510/21 and 550/30 bandpass filters and the 525LP dichroic were purchased from Omega Optics (Austin, TX), all other filters were included in the basic configurations of LSRII and FACSAria, respectively.
All data were recorded with the FACSDiva software (version 6.1.2, Becton Dickinson) and further processed by FlowJo software (version 8.8.4, Tree Star, Ashland, OR) in the form of list-mode data files version FCS 3.00 using biexponential transformation. Data can be obtained from the corresponding author. Cytometry experiments were performed in May and June, 2009. For the detection of SSEA-1, cells were stained for 45 min with a PE-conjugated anti-SSEA-1 antibody (monoclonal mouse-IgM, clone MC-480, R&D Systems, Minneapolis, MN).
Fibroblasts were seeded at 3,000 cells per cm2 and exposed to a transduction solution containing the viral vectors, 2 mM valproic acid (Sanofi Aventis, Paris, France), and 10 μg mL−1 protamine sulfate (Sigma) in MEF medium for 24 h. Subsequently, culture conditions were switched to mouse iPSC or human iPSC conditions, respectively. In the sorting approach, the sorted fractions were seeded on inactivated C3H MEF feeders, which were also used for the culture of obtained subclones. The unequally transduced murine fibroblasts were not passaged between transduction and AP staining. Human fibroblasts were seeded in lesser density (1:3) on irradiated CF1 MEF feeders 5 days after transduction.
Alkaline Phosphatase Assay
Cultures were fixed for 1 min with 4% formaldehyde (Merck, Darmstadt, Germany), permeabilized with 0.1% Tween 20 and 0.1% Triton X 100 (Roth, Karlsruhe, Germany) for 5 min, and stained (Alkaline Phosphatase Detection Kit, Millipore, Billerica, MA). Positive colonies were counted twice, using a supportive grid device.
Total RNA was isolated with the peqGold Total RNA Kit (Peqlab, Erlangen, Germany). DNA was digested using the RNase-Free DNase Set (Qiagen, Hilden, Germany). 2 μg RNA per sample was reverse transcribed with SuperScript III First Strand (Invitrogen, Carlsbad, CA) and random hexamer primers. Quantitative PCR was performed on a StepOnePlus Real-Time PCR platform (Applied Biosystems, Foster City, CA). The samples contained 2 μL cDNA preparation, 0.75 μL TaqMan probe, and TaqMan Universal PCR Master Mix (Applied Biosystems) in a total volume of 15 μL. TaqMan probes were purchased from Applied Biosystems.
Descriptions of additional methods can be found in the Supplementary Information.
Quantifiable Reprogramming Vectors Allow Cell Sorting According to Factor Stoichiometry
To trace the expression levels of the individual RFs in infected cells, we used four bicistronic lentiviral vectors ( 8) coexpressing fluorescent reporters (eGFP, dTomato, Cerulean, and Venus) along with the canonical RFs (Oct4, Sox2, Klf4, and c-Myc, Fig. 1A). We then analyzed mean fluorescent intensities after transduction of murine embryonic fibroblasts (MEFs) with a range of viral stocks' dilutions (1:100 to 1:2,500), covering a 25-fold difference in the number of infectious virus particles. In these experiments, we found that the mean fluorescence intensity of infected cell populations was indeed dependent on applied viral amounts in a linear manner (Fig. 1B). These findings allowed us to use the reporter fluorescence as a proportional measure for actively integrated vectors in individual cells.
Even more important, applying electronic compensation, we were able to independently quantify each fluorophore and could thus measure fluorescence intensities of all four fluorophores separately (Fig. 2). Hereby, eGFP, dTomato, and Venus could be easily distinguished from untransduced cells, but Cerulean expression was rather dim and the positive cells were more difficult to distinguish from untransduced cells.
To produce cell populations with different and well-defined transgene expression profiles, we transduced a bulk of MEFs with equal amounts of all four viruses [multiplicity of infection (MOI) = 4 each], resulting in a mixture of single fluorophore-positive and multifluorophore-positive cells (Fig. 3A). Next, we sorted these cells 5 days after transduction according to the respective fluorescence intensities, which reflected the individual RF expression levels. Subpopulations for all combinations of high or moderate expression of the four factors were defined by suitable gating (Fig. 3B). Reanalyses of the subpopulations, as shown for the Oct4high, Sox2low, Klf4low, and c-Mychigh fraction (Fig. 3C), could confirm the success of our sorting strategy. In the following, relative levels of the four factors will be referred to as O/o (Oct4high/Oct4low), S/s (Sox2high/Sox2low), K/k (Klf4high/Klf4low), and M/m (c-Mychigh/c-Myclow). Samples of 2,200–27,000 cells for 15 of the 16 possible combinations were initially seeded on feeder layers (Table 1). Only the fraction oSkm (Oct4low, Sox2high, Klf4low, and c-Myclow) was too slight to be used in further experiments.
|RF stoichiometry||% of total four-fold positive||Number of collected cells|
Numbers of AP-Positive and ESC-Like Colonies Differ Between Sorted Fractions
One of the early hallmarks of successful reprogramming by expression of the four RFs is a compaction of the transduced cells and a growth in colonies that acquired the characteristics of embryonic stem cells (ESCs). One of the earliest markers during this process is the gain of alkaline phosphatase (AP) activity. Remarkably, only seven of the 15 populations gave rise to AP expressing cells (Fig. 4A), most of them belonging to Oct4high stoichiometries. In contrast, elevated levels of Sox2 and/or Klf4 completely impeded colony formation when combined with low Oct4 dosages. We observed considerable differences in reprogramming efficiency, as the two fractions with Oct4 in excess over both Sox2 and Klf4 (Oskm and OskM) produced much more AP-positive and especially ESC-like colonies. In this combination, high or low expression of c-Myc had apparently little influence (Fig. 4B). It is noteworthy to mention that besides colonies consisting of fully reprogrammed iPSCs, which completely silenced the exogenous RFs expression, we also observed some partially reprogrammed colonies, which did not fully acquire the features of pluripotency and which had not silenced the expression of RFs, as visualized by residual reporter fluorescence before AP staining (Supporting Information Fig. 1).
As these results suggested that the subpopulations characterized by high transcripts levels of Oct4 are superior to those which had received equimolar amounts of the four factors, we validated our approach by intentionally transducing the same murine fibroblasts that were used in the initial experiment with nonequimolar RF stoichiometries. To this end, we applied low, medium, and high viral amounts by choosing MOIs for O:S:K:M of 1:1:1:1, 2:2:2:2, and 3:3:3:3 for the equimolar stoichiometries and 2:1:1:1, 4:2:2:2, and 6:3:3:3 for the Oct4high stoichiometries. Because the single factors were coupled to a fluorophore, we could confirm the actual obtained expression levels by flow cytometry (Supporting Information Fig. 2A), and we were able to demonstrate that we indeed achieved the intended stoichiometric ratios. By further cultivation of these samples and by subsequently assessing the number of AP-positive colonies and those which in addition exhibited ESC-like morphology, we support our findings that Oct4high stoichiometries resulted in more efficient reprogramming of MEFs into iPSCs independently from the application of low, medium, or high total amounts of RFs (Supporting Information Fig. 2B). When we did a similar experiment for reprogramming of human fibroblasts, we lowered the viral amounts of O:S:K:M to MOIs of 0.5:0.5:0.5:0.5 and 1:1:1:1 or 1:0.5:0.5:0.5 and 2:1:1:1 for the equimolar and nonequimolar stoichiometries, because human fibroblasts are more permissive for the VSV-G pseudotyped viral particles. Again, we obtained more AP-positive colonies in the Oct4high stoichiometries, and we confirmed the dose-dependent reprogramming efficiency according to the total amount of applied RFs also in human cells (Supporting Information Fig. 3).
Molecular Characteristics of iPSCs Derived from the Sorted Subpopulations
Interestingly, iPSC clones obtained from different cell sorting fractions showed some notable dissimilarities: while three out of the five cell lines featured a uniform compact, sharply bordered ESC-like morphology and complete silencing of the exogenous reprogramming genes over further passages, two cell lines, namely OsKm and OSKm, displayed residual reporter fluorescence along with a loosened and disaggregated colony morphology (Fig. 5A). To investigate molecular differences related to RF stoichiometry, we first analyzed expression of the stage-specific embryonic antigen 1 (SSEA-1) by flow cytometry. The three lines with intact ESC-like morphology (Oskm, OskM, and OsKM) showed explicit SSEA-1 reactivity with frequencies of 68.7%, 84.0%, and 79.2%, respectively (Fig. 5B). Because the iPSCs were cultivated on feeder cells and the separation of remaining MEFs was incomplete, some cells showed the same background reactivity as SSEA-1-stained MEFs (red curves in Fig. 5B). In contrast, the two less perfect iPS cell lines (OsKm and OSKm) showed only few SSEA-1-positive cells (27.2% and 12.1%, respectively). Of note, the cell preparations contained cells that were even less intensely stained than the faintly SSEA-1-expressing control fibroblasts, indicating that derivatives of these lines might have been differentiated away from the pluripotent stage toward a more differentiated SSEA-1-negative stage. Besides surface antigen characterization, we performed quantitative RT-PCR for gene expression analyses of the major pluripotency-associated transcription factors Oct4, Sox2, and Nanog as well as transcript levels of Gtl2 and Rian, both of which are encoded in the maternally imprinted Dlk1-Dio3 locus on chromosome 12, and which were suggested to closely reflect the developmental capabilities of murine iPSCs ( 10). Remarkably, the expression profiles of these genes and transcripts were in general quite similar in all five established iPSC lines (Fig. 5C), with the exception that iPSCs derived from the OSKm-sorted fraction expressed endogenous Oct4 and Sox2 as well as Rian at manifestly substandard levels.
Finally, we assessed the in vitro differentiation capabilities of the five iPSC lines by aggregating the cells in hanging drops to form embryoid bodies (EBs) and by applying cell culture conditions permissive for spontaneous differentiation of pluripotent stem cells. All iPSC lines gave rise to derivatives of all three germ layers within the EBs, as indicated by qRT-PCR detection of transcripts (Supporting Information Fig. 4A) representing endoderm (GATA-binding protein 4 [Gata4] and α-fetoprotein [Afp]), mesoderm (Brachyury [T] and VEGF-receptor [Kdr]), ectoderm (keratinocyte transcription factor [Trp63] and orthodenticle homeobox 2 [Otx2]), as well as the germ-line (Stella [Dppa3]). The gene expression data were further substantiated by plating EBs (Supporting Information Fig. 4B) on gelatin-coated dishes and by applying medium conditions supportive for neural differentiation (tubulin β3, Supporting Information Fig. 4C) for cardiomyocyte differentiation (cardiac troponin T, Supporting Information Fig. 4D) and for hepatic differentiation (α-fetoprotein, Supporting Information Fig. 4E).
Human ESC research originated great promise in regenerative medicine because of the cells' strong self-renewal capacity and their ability to differentiate into functional derivatives of all three germ layers. The seminal discovery of an approach for direct reprogramming of somatic cells into pluripotent cells by retrovirus-mediated expression of pluripotency associated genes by Takahashi and Yamanaka in 2006 ( 1) resulted in a vast increase in pluripotent stem cell research for both regenerative stem cell-based therapies (11, 12) and research on disease-specific induced pluripotent stem cells (13). The generation of human iPS cells (13–15) together with the latest development showing production of iPS cells without integrating vectors (16–19) create new opportunities for the establishment of clinically useful autologous stem cell lines. However, although a lot of progress has been described for further applications of induced pluripotent stem cells with respect to safety aspects avoiding transgene integration of the RFs, still little is known about the optimal stoichiometric ratio of the individual RFs. In only one study, the iPSC generation efficiencies were compared, when different loads of lentiviral vectors were applied to the starting human fibroblasts (6). However, taking into account that the iPSC generation process is a quite rare event and usually less than 1% of the transduced cells successfully undergo reprogramming, the approach of comparing high vs. low amounts of RF encoding viruses in a limited cell number might disregard random effects because of differences in lentiviral integration sites and copy numbers among various cells. For instance, it could not be demonstrated whether the high amounts of Oct4-expressing viruses also led to an evenly high-Oct4 expression in all transduced cells. To circumvent these issues, we established a multicolor cytometry based sorting approach that enabled us to quantify moderate vs. high expression levels of four distinguishable fluorophores (eGFP, dTomato, Cerulean, and Venus) and modified previously published filter settings to discriminate the eGFP and Venus (YFP) signals (20, 21) while using a violet laser for Cerulean detection (22). We chose these fluorophores, because transgenic expression of the respective cDNAs is achievable in cells without major influences on cell viability or other biological properties (7, 23) and constructed lentiviral vectors allowing IRES-mediated coexpression of specific fluorophores with each single RF (Oct4-eGFP, Sox2-dTomato, Klf4-Cerulean, and c-Myc-Venus) based on our recently described modular lentiviral expression system for the generation of iPS cells (8). This validated lentiviral system ensures high-transduction efficiencies and transgene expression levels and thus allowed us to obtain reasonable numbers of multifluorescent fibroblasts for our sorting approach. Only the subpopulation Oct4low, Sox2high, Klf4low, and c-Myclow was too tiny to result in a cell number usable for further experiments. This might be due to a necessary gating strategy compromise for the Cerulean and dTomato channels, which was caused by the rather dim Cerulean signals. Nevertheless, reanalyses confirmed the suitability of our defined gates to distinguish the stoichiometric ratios of the four RFs in the various subpopulations.
In agreement with a study from Papapetrou et al. ( 6), our findings suggest that increased Oct4 levels opposite to modest ones for Sox2 and Klf4 are required for satisfying reprogramming efficiencies. In our study, we provide additional evidence that high Oct4 levels are crucial for achieving a stable pluripotent state independent of ectopic expression of RFs, which are supposed to be silenced upon successful and complete reprogramming (8, 24). In extension to the previous reports, we also included those stoichiometries with two highly expressed RFs and two moderately expressed RFs in our analyses and were able to demonstrate that besides high Oct4 expression alone, also combined Oct4high and c-Mychigh expression is beneficial for efficient generation of fully pluripotent iPSC lines. Interestingly, high levels of Sox2 and/or Klf4 appeared to have a pronounced detrimental effect on iPSC generation efficiency, when combined with high Oct4 levels. These findings are in line with previous studies reporting that Oct4 is the most essential pluripotency factor, and that ectopic expression of Oct4 alone is sufficient to reprogram distinct cell types (5, 25). Interestingly, reprogramming is still possible, but delayed and less efficient, in the absence of ectopic c-Myc expression. It seems, however, that c-Myc greatly enhances the initial steps of reprogramming by downregulation of fibroblast-specific genes at the beginning of the reprogramming process ( 26). In this context, it is noteworthy that nonintegrating episomal vectors were successfully used to generate iPSCs by overexpressing Oct4 alone in neural stem cells that constitutively expressed c-Myc (27).
In previous studies, it was suggested to omit c-Myc from the RF cocktail, because the role of c-Myc in reprogramming was considered to be a double-edged sword: it was reported not only to be primarily beneficial for the number of colonies but also to impair the generation of fully reprogrammed iPSCs ( 3, 28). In contrast, we demonstrate that the molecular characteristics measured by pluripotency associated factors such as Oct4, Sox2, and Nanog gene expression and SSEA-1 surface protein expression is unaltered in the OskM subpopulation when compared with the Oskm subpopulation.
Our results suggest the necessity to address stoichiometry also in alternative approaches for iPSC generation such as use of nonintegrating viral vectors ( 17, 18, 29, 30), protein transduction (31, 32), plasmid transfection (16), or stabilized mRNA application (33). The importance of our findings is underlined by the fact that in the latter case, Warren et al. (33) used elevated Oct4 levels and achieved a highly efficient iPSC conversion, while all other nonintegrating approaches so far reported extremely low efficiencies.
In conclusion, applying a multicolor flow cytometric sorting approach, we could provide further evidence on the impact of high Oct4 expression during the course of reprogramming. Most importantly, we demonstrated that the eligible Oct4high, Sox2low, and Klf4low subpopulations only resemble a small fraction of cells targeted by equal vector amounts. Finally, with our data, we emphasize that flow cytrometry based stoichiometric analyses will be of particular importance if efficiencies are compared between different experimental systems.
The authors are grateful to Françoise André for the technical support and to Christopher Baum for the helpful comments and insightful discussions.