Dissecting T cell lineage relationships by cellular barcoding

To uniquely tag individual cells with molecular barcodes, we generated a retroviral plasmid library, consisting of ∼5,000 different plasmids. Each plasmid contains GFP as a marker gene and a so-called barcode, formed by a semirandom stretch of 98 bp of noncoding DNA. To allow for rapid analysis of barcodes present within a given cell population, each barcode present in this library was also amplified individually by PCR and spotted on microarray slides, resulting in a barcode microarray. In a typical experiment, cells are labeled by retroviral transduction with the barcode library and reintroduced into mice. After in vivo expansion and differentiation of the barcode-labeled cells, barcodes present in different subsets of cells are recovered by PCR, fluorescently labeled, and hybridized on the barcode microarray. Determining the overlap of barcodes between different subsets by cellular barcoding (Fig. 1 A) should then allow one to determine the kinship between these cell populations. In addition to the retroviral barcode library used here, a lentiviral barcode library was prepared to allow tagging of quiescent cell types that are refractory to retroviral infection (unpublished data).

Before assessing the feasibility of kinship analysis, we first established whether the barcode microarray shows any crosshybridization between different barcodes present in the barcode library. When the barcode microarray was hybridized with PCR products of 273 individual Escherichia coli clones of the master library that were labeled with Cyanine-3 (Cy3), with Cy5, or with both Cy3 and Cy5 dyes, no detectable crosshybridization between different barcodes was observed (Fig. 1 B). Second, as the effective size of the barcode library will decrease proportionally with the number of barcodes that is introduced per cell, we also established retroviral transduction conditions under which the mean number of barcodes per cell is close to one (Fig. S1).

As a first test of the barcoding technology, we assessed whether cellular barcoding can be used to analyze kinship between cell populations that are either kin or nonkin. For this purpose, OVA-specific MHC class I–restricted T cell receptor transgenic OT-I T cells were transduced with the barcode library. Subsequently, sorted barcode-labeled OT-I T cells were i.v. injected into mice that were then challenged with antigen (i.v. challenge with Listeria monocytogenes encoding OVA [LM-OVA]). At day 7 after challenge, the OT-I T cell pool present in the spleen was split into two samples (which are evidently derived from the same pool of precursors), hereafter referred to as half-samples. From the genomic DNA of these half-samples, barcodes were recovered by PCR, labeled with Cy3 and Cy5 dyes, and cohybridized on the barcode microarray. Resulting data are presented in two-dimensional plots, depicting the Cy3 and Cy5 signals of each individual spot on the barcode microarray. In these plots, each dot represents a clone (or family) of cells that are the progeny of a precursor containing a specific barcode. The presence or absence of a barcode within a cell population is then determined by calculation of the distribution of background signals from averaged Cy3 and Cy5 signals, as described in Materials and methods. When two half-samples from the same spleen are compared for the presence of barcodes, >85% of the barcodes that show detectable signal in one of the spleen samples is also detected in the other half-sample (Fig. 2 A, left and middle). In contrast, when barcodes recovered from spleen samples of two different mice (representing two cell pools that are nonkin) are compared, the barcodes that are detected are largely unique to either one of the spleen samples (Fig. 2 A, right). The small percentage of the barcodes that is shared between the T cell populations in the two mice is close to the overlap that is expected based on chance (i.e., the introduction and activation of T cells that by chance have the same barcode in the different mice; Fig. 2 A, legend). Together these data indicate that cellular barcoding can distinguish between cells that are kin or nonkin and establish the patterns of barcode overlap that may be expected under conditions of full kinship or lack thereof.

Subsequently, we determined whether cellular barcoding can also detect more subtle differences in kinship, reflecting situations where only part of the T cells in two effector T cell pools has a common ancestor in the naive T cell compartment or where the progeny of a given naive T cell is enriched but not unique to a certain T cell population. For this purpose, we performed microarray hybridizations of barcodes isolated from genomic DNA of two spleen samples that were mixed at different ratios (1:1, 3:1, 7:1, 15:1, and 1:0). In the most extreme situation, where part of the barcodes is unique to one of the samples and the remainder is shared (i.e., the 1:1 versus 1:0 comparison; Fig. 2 B, right), those barcodes that are common or private (represented by green and red dots, respectively) are readily distinguished. Furthermore, in situations where part of the barcodes is overrepresented 2–8-fold in one of the samples, these barcodes (Fig. 2 B, red dots) become progressively separated from the other barcodes. In the analysis shown in Fig. 2 B, relatedness of cell populations is revealed by grouping barcodes as either shared or private (i.e., present in either the top right quadrant or the bottom right/top left quadrant). As an alternative measure of relatedness, the correlation between the Cy3 and Cy5 hybridization signals for each barcode can be used. Analysis of the data in Fig. 2 B by such correlation analysis demonstrates that an increasing difference in the frequency of half of the barcodes in two samples leads to a progressive loss in correlation (Fig. 2 B, r values). These data indicate that cellular barcoding can be used to reveal cellular progeny that are enriched in a certain cell population and demonstrate that the fluorescence intensity of barcode microarray signals correlates with the frequency at which a specific barcode is present within a certain sample.

Having established the feasibility of lineage analysis by cellular barcoding, we assessed whether T cell families that accumulate at distinct priming and effector sites share common ancestors in the naive T cell pool. It has previously been shown that T cell subsets can be distinguished on the basis of differential expression of chemokine receptors and adhesion molecules, and such differential expression is thought to guide the migration of these subsets to distinct sites in the body (5–9). For instance, it has been shown that T cells that are activated in vitro by dendritic cells from gut draining LN (DLN) or skin DLN preferentially migrate to the corresponding effector site (10–12) upon reintroduction. To determine whether selective migration of T cells also occurs upon antigen encounter in vivo in a case where there is a choice between two antigen-positive effector sites, we analyzed the kinship between antigen-specific T cells in a dual challenge model. For this purpose, mice were challenged by s.c. inoculation of EL4-OVA tumor cells and parallel intranasal infection with WSN-OVA influenza virus to generate two local independent effector sites. We considered the possibility that a potential effect of imprinting could be masked or reduced when T cell activation in the two sets of DLN would not occur simultaneously as, in this case, the progeny of T cells activated within the first set of DLN could seed the other LN bed (Fig. 3) (13) and thereby dominate the T cell response at that site. In particular, in view of recent data that suggest that programming of T cell functions may not be fixed (10, 12), such a situation could lead to an underestimate of the potential of imprinting. To avoid this issue, we first determined the kinetics of T cell priming at the two different LN sites. Analysis of CFSE dilution of labeled OT-I T cells in tumor and lung DLN of mice that received an s.c. inoculation of EL4-OVA tumor cells or intranasal WSN-OVA influenza infection revealed that detectable T cell proliferation is first observed at days 3 and 4 after antigenic challenge, respectively (unpublished data). To allow simultaneous priming of T cells in the two LN beds, mice were therefore challenged with EL4-OVA tumor inoculation 1 d after intranasal WSN-OVA influenza inoculation in all subsequent experiments (Fig. 3).

The kinship of antigen-specific T cells that accumulate at the two priming and two effector sites was subsequently monitored by barcode analysis. When barcodes are recovered from tumor DLN at day 5 or later time points, the percentage of barcodes recovered from two half-samples of a tumor DLN that is found in both half-samples is markedly higher than the random overlap observed between tumor DLN samples obtained from different mice (Fig. 4). Likewise, when barcodes are recovered from lung DLN, the percentage of barcodes recovered from two half-samples of a lung DLN that is found in both half-samples is markedly higher than the background overlap observed between lung DLN samples from different mice (Fig. 4). Together, these data indicate that barcode analysis of LN-resident T cell populations is feasible, even early after the induction of T cell proliferation.

To reveal whether draining LNs of tumor and lung contain the same sets of T cell families, barcodes obtained from tumor DLN and lung DLN were subsequently compared. Early after antigenic challenge (day 5), the overlap in barcodes present in tumor and lung DLN obtained from the same mouse (Fig. 4, A [right, top two rows] and C) is small and only slightly higher than the overlap between barcodes obtained from tumor and lung DLN from different mice (Fig. 4, A [right, bottom row] and C). This indicates that early after antigen-induced activation, T cell families mostly remain confined to the original site of priming. Interestingly, from days 6 to 8 after challenge, a marked shift occurs, in which a large fraction of barcodes becomes shared between T cells isolated from tumor and lung DLN, suggesting that within days after priming, the offspring of T cells intermingles within different DLN.

To provide a more quantitative assessment of the similarity of T cell families present at the two sites, data were converted into percentage identity, in which the percentage of shared barcodes, as shown in Fig. 4 C, is corrected for maximal and background overlap (analogous to the correction of data for background and maximal lysis in CTL assays, see Materials and methods for details). Based on this analysis, T cell families within tumor and lung DLN share 8 and 16% identity on days 5 and 6. This abruptly changes to 78 and 75% identity on days 8 and 10. Also, when the data are analyzed by correlation analysis, T cell families at the two sites are clearly distinct early after antigenic challenge but become close to identical from day 8 (Fig. S2). To test whether the increase in barcode overlap between tumor and lung DLN over time is indeed caused by T cell migration, we treated mice with FTY720, a compound which blocks LN egress of T lymphocytes (14, 15). Importantly, when LN egress is blocked during the course of antigenic challenge, the relatedness of T cells that are present in tumor and lung DLN remains low even late during infection (day 8, 29% identity; Fig. S3). These data show that the time-dependent increase in relatedness of T cell families isolated from lung and tumor DLN is a result of T cell migration.

To test whether barcode analysis of effector site–resident T cells was feasible, barcodes were obtained from tumor and lung tissue at various time points after antigenic challenge. When T cells are isolated from the tumor and lung effector sites during the OVA-specific T cell response, barcodes can be recovered reliably from these samples (Fig. 5, A [left and middle, top two rows] and B). Importantly, of the barcodes that are present at one effector site, the majority is also found at the other effector site within the same mouse (Fig. 5, A [right, top two rows] and B), whereas samples derived from different mice do not show such a shared set of barcodes (Fig. 5, A [bottom row] and B). Furthermore, this high overlap is observed both at day 8, when substantial T cell accumulation at both sites has just begun (Fig. 5), and at the peak of the antigen-specific T cell response (day 10 after challenge; not depicted). Thus, although the T cell populations that develop within tumor and lung DLN initially consist of distinct families, the progeny that accumulates at the lung and tumor effector sites are highly related (81% identity at day 8; Fig. S4 A) Together, these data show that, in this model, the progeny of CD8⁺ T cells disperses, essentially indiscriminately, over two effector sites.

To investigate whether this pattern also applies to a third effector site, we analyzed the kinship of T cells migrating to inflamed gut and s.c. tumor effector sites. To this purpose, mice were challenged by oral infection with LM-OVA and s.c. inoculation of EL-4 OVA tumor cells. Again, challenges were timed such that priming occurred simultaneously in gut and TDLNs (unpublished data). Furthermore, by comparing barcodes isolated from different half-samples around the peak of the OVA-specific T cell response, it was established that barcodes from gut tissue can also be recovered reliably (Fig. 6). When barcodes present in T cells that accumulated in the gut and the tumor of the same mouse were compared, no evidence was found for selective migration of T cell families toward the tumor site (Fig. 6, A [right, top two rows] and B; and Fig. S4 B). In some mice, a small fraction of barcodes was enriched in gut tissue, suggesting that selective migration toward the gut may occur to some degree. However, the majority of T cell families also appeared to be equally represented at these two effector sites (translating into an identity of 64% between T cell families in gut and tumor of the same mouse). As expected, tissue samples derived from different mice did not show such a shared set of barcodes (Fig. 6, A and B, bottom rows). Together, these data indicate that also in this comparison, the majority of T cell families that locally develop accumulate at both effector sites.

Finally, a possible limitation of the experiments described in the previous paragraphs is that the 3-d in vitro culture used to introduce the Moloney-based retroviral barcode library may conceivably influence T cell migration properties. To address this issue, we have developed a strategy in which naive barcode-labeled T cells are obtained via the transduction of OT-I thymocytes (see Materials and methods for details). Also, when naive barcode-labeled OT-I T cells were used, a high degree of overlap between barcodes present in T cells that accumulated in the lung and the tumor of the same mouse was observed (Fig. S5, A [right, top two rows] and B).

In this paper, we developed a novel approach, termed cellular barcoding, to dissect lineage relationships between different cell populations. In this approach, retroviral or lentiviral tagging of individual cells with molecular barcodes is coupled to a microarray-based detection system to allow kinship analysis of the progeny of such cell populations. By analyzing family relationships between different barcode-labeled T cell populations, we demonstrate that cellular barcoding can be used to distinguish cells that are kin or nonkin. In essence, this technology provides an in vivo system for high throughput parallel clonal assays.

To exemplify the potential of cellular barcoding, we analyzed the kinship of CD8⁺ T cells that accumulate at different DLN and effector sites. A picture that has emerged from recent studies is that T cells primed within a specific LN, or by antigen-presenting cells obtained from LNs draining a specific tissue site, acquire a phenotype that is consistent with preferential migration (10–13, 16, 17). Furthermore, cells that are induced to display such a phenotype in vitro also show selective homing upon injection (11, 12). Interestingly, one study showed that the in vivo expression of markers that could result in preferential migration is transient, possibly through the action of signals encountered in other LN beds (13). In addition, this reprogramming has also been observed in vitro (10, 12).

In the current paper, we have used cellular barcoding to study the kinship between T cells that accumulate at different effector sites in situations where antigen is available at multiple locations. Data obtained in this analysis demonstrate that from the first time point at which substantial amounts of T cells accumulate at the tumor and lung effector sites, the T cell families at both sites are largely identical, even though unique T cell families are found in the LNs draining these tissues shortly after antigen challenge. In addition, a high identity was also observed when comparing T cell families that accumulate in inflamed gut and s.c. tumor tissue. In this case, a small set of barcodes appears enriched in gut tissue in at least some animals. Although these data suggest that a minority of T cell families might display the capacity to selectively migrate to this site, based on signal intensity, the contribution of this set of clones to the total T cell pool in gut is not expected to be large. Collectively, these data indicate that after a brief period of local accumulation, the dominant pattern of T cell accumulation at effector sites is aselective. We speculate that selective T cell migration may only play a significant role at very early time points during infection (13) and perhaps in situations where the effector sites generate limited inflammatory signals. In contrast, we suggest that aselective migration forms the dominant process at the moment two effector sites produce sufficient proinflammatory signals, and aselective migration, in fact, seems a preferred strategy to combat spreading infections.

Interestingly, a strong time-dependent increase in the relatedness of T cell families isolated from lung DLN and tumor DLN of the same mouse was observed. Furthermore, this increase was found to be dependent on lymphocyte egress from the LN, as it was blocked in the presence of FTY720. The mingling of T cell families at two LN sites is consistent with the LN recycling of recently activated T cells observed by Liu et al. (13) but also with the recent description of effector T cell entry into reactive LNs via high endothelial venules (18). As a result of such T cell redistribution over tumor and lung DLN, an indiscriminate accumulation of T cell families at the two effector sites may occur as a secondary effect (Fig. 7 A). As an alternative model, accumulation of effector T cells at both sites could occur independently of LN cross-seeding (Fig. 7 B), and equilibration of LN T cell populations could then occur as a secondary effect, through T cell recycling from effector sites via the afferent lymph (19). In view of the rapid equilibration of T cell families in lung and tumor DLN between days 6 and 8, a time point at which T cell numbers at effector sites are still low, we favor the former possibility, but formal evidence is lacking.

Together, the current data demonstrate that cellular barcoding can be used to dissect the migration patterns of T cell families in vivo and show that the majority of T cells that are activated during a T cell response acquire the capacity to accumulate at multiple tissue sites. In addition to the potential value of the barcoding technology for the analysis of cellular differentiation and migration within the lymphoid system, we consider it likely that this technology can also be used to examine stem cell, cancer stem cell, and differentiation issues for other hematopoietic and transplantable cell types.

C57BL/6 (B6), B6 Ly5.1⁺ and C57BL/6.OT-I mice (OT-I) (20) were obtained from the animal department of the Netherlands Cancer Institute. All animal experiments were performed in accordance with institutional and national guidelines and were approved by the Experimental Animal Committee (DEC) of the Netherlands Cancer Institute.

The semirandom stretch ((N)₈-(SW)₅)₅-N₈ was inserted 3′ of the GFP ORF in the pLentilox3.4 vector (Supplemental Materials and methods). Subsequently, 4,743 individual E. coli cell clones containing this vector were picked and grown in 96-well plates. This master library was used to generate a barcode microarray and a Moloney-based retroviral library (supplemental Materials and methods). To generate the barcode microarray, PCR products were generated from the individual E. coli cell clones of the master library and were spotted in duplicate onto poly-l-lysine–coated glass slides using the MicroGrid II arrayer (Isogen Life Science). PCR products of 93 additional E. coli cell clones that are not present in the master library were spotted in hexaplicate to function as internal controls for the labeling and hybridization procedure (supplemental Materials and methods).

Retroviral supernatants were generated using Phoenix-E packaging cells, as described previously (21), and stored at −80°C. Total OT-I splenocytes were isolated and mixed with CD4-enriched B6 splenocytes at a 10:1 ratio. CD4-enriched B6 splenocytes were obtained by staining B6 splenocytes with PE- or APC-conjugated anti-CD4 mAbs (BD Biosciences), labeling with anti-PE or anti-APC Microbeads (Miltenyi Biotec), followed by Midi-MACS enrichment (Miltenyi Biotec). Enrichment was ≥90%. Cells were activated in IMDM (Invitrogen) supplemented with 8% heat-inactivated FCS, 100 U/ml penicillin, 100 μg/ml streptomycin (Boehringer Ingelheim), and 0.5 × 10⁻⁵ M β-mercaptoethanol (culture medium) in the presence of ConA and IL-7 and transduced with diluted retroviral supernatants by spin infection as previously described (21). Retroviral supernatants were used at dilutions resulting in transduction efficiencies of ∼1–2%. At this transduction efficiency, the mean number of barcodes per cell is close to one (Fig. S1). 24 h after retroviral transduction, transduced T cells were purified using Ficoll Paque PLUS (GE Healthcare), stained with PE-conjugated anti-Vα2 and APC-conjugated anti-CD8α, and sorted on a FACSAria (BD Biosciences) for GFP⁺ Vα2⁺ CD8α⁺ cells. Subsequently, the cells were washed once in culture medium and twice in HBSS (Invitrogen), resuspended in HBSS, and injected i.v. (1,000–10,000 GFP⁺ cells per mouse). To obtain naive barcode-labeled T cells, OT-I thymocytes were transduced with the barcode library in culture medium supplemented with 10 ng/ml IL-7. The next day, transduced cells were sorted for GFP expression and injected in the thymi of B6 recipient mice (∼10⁶ cells/mouse). 3 wk after intrathymic injection, CD8⁺ T cells containing barcode-labeled OT-I cells were isolated from spleen and LNs. Barcode-modified OT-I T cells generated in this manner are phenotypically naive and respond to antigen encounter with identical kinetics as unmanipulated OT-I T cells (unpublished data).

The recombinant influenza A strain WSN-OVA that expresses the H-2K^b–restricted OVA_257-264 epitope (SIINFEKL) has been previously described (22). A recombinant influenza A virus that expresses epitope IV of the SV40 large T antigen was generated as previously described (22). Influenza A strains were grown in and titered on MDCK (Madin Darby canine kidney) cells. Mice were infected with 500 PFU of the influenza A variants by intranasal application. EL4-OVA tumor cells were produced by retroviral transduction of EL4 thymoma cells with a pMX vector that encodes a GFP-OVA_257-264 fusion protein (23), and EL4-NP cells have been previously described (24). For tumor challenge, tumor cells were washed three times with HBSS, resuspended in HBSS, and injected s.c. into mice (10⁷ cells per mouse). The LM-OVA recombinant L. monocytogenes strain that expresses OVA (25) was provided by D. Busch (Technical University Munich, Munich, Germany). Mice were infected by gavage with ±10⁸ CFU LM-OVA or i.v. with ±2.5 × 10⁴ CFU LM-OVA.

Cell suspensions of spleen, LN, lung, and tumor tissue were generated as previously described (26). Cell suspensions of gut tissue were generated according to a previously published procedure (24). In brief, the small intestine was flushed with 50 ml HBSS containing 2% FCS. Peyers patches were excised and intestine was cut into small pieces longitudinally and laterally. After washing, the tissue was transferred to a 50-ml tube containing 20 ml HBSS/5% FCS/0.1mM EDTA and stirred for 20 min at 37°C. The supernatant was harvested, EDTA treatment was repeated twice, and all supernatants were pooled. Subsequently, the material was digested in 20 ml IMDM containing 200 U/ml collagenase (type VIII; Sigma-Aldrich) and stirred for 1 h at 37°C. The supernatant was harvested and collagenase treatment was repeated once. Supernatants were pooled and the thus-obtained cell suspension was used for subsequent enrichment of OT-I cells. To enrich OT-I cells from tissue samples, cells were stained with PE-conjugated anti-Vα2 mAbs (BD Biosciences), labeled with anti-PE Microbeads (Miltenyi Biotec), followed by autoMACS (Miltenyi Biotec) enrichment. For analysis of barcodes from T cells present in LNs, LN cells were first expanded for 4 d in culture medium supplemented with 20 U/ml of recombinant human IL-2 (Chiron) and 12 ng/ml of recombinant mouse IL-7 (PeproTech). Half-samples were cultured separately to avoid bias introduction during in vitro culture. The mAbs used for flow cytometric analysis were APC-conjugated anti-CD8α (BD Biosciences) and PE-conjugated anti-Ly5.2 (eBioscience). Before analysis, 1 μg/ml propidium iodide (Sigma-Aldrich) was added to enable selection for propidium iodide–negative (living) cells.

Genomic DNA was isolated using a DNeasy tissue kit (QIAGEN) and barcodes were amplified by nested PCR (supplemental Materials and methods). PCR products were purified, concentrated, labeled with Cy3 or Cy5 fluorescent groups, and hybridized to the barcode microarray (supplemental Materials and methods). Fluorescence intensities, as quantified using Imagene 6.0 (BioDiscovery, Inc.), were corrected for background noise and normalized according to Yang et al. (27).

After averaging of duplicates, barcodes present above background were selected based on the probability that a signal differed from an artificial background distribution. In barcode microarray hybridizations, the majority of signals reflects a negative biological signal and, consequently, the mode of the signal distribution can be taken as an approximation of the mean of the background distribution. As signals below the mode are the least contaminated by positive signals, an artificial background distribution was constructed from the signals lower than the mode and an extrapolation/mirroring of these signals to the positive side of the slope. For each sample, this artificial background was determined based on the distribution of the averaged Cy3 and Cy5 signals of the hybridization in which two half-samples of the same tissue were hybridized (the cutoffs used [p-values given in the figure legends] are indicated by the horizontal and vertical dividers in the dot plots). The two dimensional plots depict Cy3 and Cy5 signals for all 4,743 barcodes that were present in the master library. For comparison of barcodes in samples analyzed in different hybridizations, we performed in silico hybridizations in which the Cy3 and Cy5 signals of the two samples were plotted against each other. Percentage identity was calculated as:

in which A−B is the experimental comparison (EXP) and A−A and B−B comparisons of two samples from a given site, in which the samples are either from the same mouse (INTRA) or from two different mice (INTER). All numbers refer to the fraction of shared barcodes of the total number of detected barcodes in both samples. Conceptually, this calculation is analogous to the calculation of percentage of specific lysis in CTL assays, with the one distinction that the maximal and minimal signals are now defined by the efficiency of barcode recovery from two sites (A−A^INTRA + B−B^INTRA and A−A^INTER + B−B^INTER, respectively), assuming that the percentage of shared barcodes within a sample is the product of two samples that have an equal efficiency of barcode recovery.

For each barcode comparison, the Pearson Correlation between Cy3 and Cy5 signals was calculated. To avoid dominance of signals from barcodes not present within a given sample in the correlation analysis, a small fraction of barcodes present in the bottom left quadrant (1/3 of the number of barcodes present in the top left, top right and bottom quadrants) was randomly drawn and incorporated in the correlation analysis. To minimize variations in correlation values introduced by the random drawing of a subset of barcodes from the bottom left, all correlation values were averaged from 100 calculations, based on 100 random drawings from the bottom left.

Fig. S1 presents additional data showing that under the conditions used, the majority of transduced cells contains a single transgene. Fig. S2 shows the redistribution of T cell families over LN beds through time as analyzed by correlation analysis. Fig. S3 shows that the redistribution of T cell families over LN beds is dependent on LN egress. Fig. S4 shows the kinship of T cells present at different effector sites as determined by correlation analysis. Fig. S5 shows the kinship of T cells present at lung and tumor effector sites when naive barcode-labeled T cells are used. A detailed description of barcode library generation, microarray generation, barcode recovery, and microarray hybridization is available as supplemental Materials and methods.

The authors thank Drs. J. Borst, H. Jacobs, and S.H. Naik for discussions and critical reading of the manuscript, Dr. N.J. Armstrong for advice on data analysis, Dr. D. Busch for the kind provision of LM-OVA, W. Brugman for microarray production, and A. Pfauth and F. van Diepen for cell sorting.

R.A. is supported by a grant from the Netherlands Organization for Scientific Research (NWO-Veni 916.56.155).

The authors have no conflicting financial interests.