Identification of the Targets of T-cell Receptor Therapeutic Agents and Cells by Use of a High-Throughput Genetic Platform

T-cell receptor (TCR)–based therapeutic cells and agents, including adoptive T cells and tumor-infiltrating lymphocytes (TIL; refs. 1, 2), TCR-engineered T cells (3), ImmTACs (4), TCR-mimic (TCRm) antibodies (5), and neoantigen vaccines (6, 7), are cancer therapeutics that target cells expressing intracellular cancer-associated proteins. These agents rely on presentation of peptides derived from cellular, viral, or phagocytosed proteins on MHC, also known as human leukocyte antigen (HLA). However, cross-reactivities of these agents with off-target cells and tissues are difficult to predict and have resulted in serious, sometimes fatal, adverse events (8, 9). In addition, identifying the antigenic targets of TILs found in tumors is time-consuming, expensive, and complicated (2).

TCR-based therapeutics are structurally similar to the TCR on CD8 T cells and thus share both their potential advantages and challenges. For instance, CD8 T cells can theoretically discern whether any MHC-I bound peptide is self, foreign, or altered-self. Yet, the number of possible MHC-I ligands that can be encoded by the 20 proteinogenic amino acids is significantly larger than the number of circulating T cells in the human body. To account for this discrepancy, TCRs are cross-reactive: most reports suggest that TCRs can recognize hundreds to thousands of distinct pMHC (10–14). Thymic selection in vivo is critical to deplete auto-reactive T cells. Some TCR-based therapeutics are made completely in vitro (e.g., phage display) and thus do not undergo negative selection for the human pMHC repertoire. Other TCR-based therapeutics are isolated from humans but subsequently modified to make them higher affinity, thus potentially introducing new cross-reactivities. As a consequence, each of these agents can be cross-reactive with HLA-presented peptides found in normal tissue (15). A prominent example is an affinity-enhanced TCR directed against an HLA-A*01:01 MAGE-A3 peptide (168-176; EVDPIGHLY), which induced lethal cardiotoxicity in two patients treated with these T cells during a phase I clinical trial. Extensive preclinical testing failed to uncover off-target reactivity; it was later discovered that an epitope derived from Titin (24337-24345; ESDPIVAQY), a structural protein highly expressed by cardiomyocytes, was cross-reactive with the MAGE-A3 TCR (8). Another TCR directed toward the MAGE-A3 peptide (112-120: KVAELVHFL) led to neuronal toxicity and death in several patients, likely due to cross-reactivity of the TCR to a peptide from the MAGE-A12 protein (112-120: KMAELVHFL; ref. 9). Hence, a major challenge to the development of safe TCR-based therapeutics is the prospective identification of off-tumor, off-target pMHC (16).

Identifying off-tumor, off-target pMHC is challenging because the complete repertoire of HLA ligands found in normal human tissue is unknown. The number of known HLA ligands in humans is rapidly expanding, with reports identifying thousands of novel presented peptides (17, 18). However, it is unclear how many presented peptides are not known and little is known about the antigens presented on critical tissues such as the nerves, eyes, heart, and lungs. Furthermore, cross-reactive pMHC are not easy to identify. Methods to identify cross-reactive targets of TCR-like molecules have been developed by testing yeast (11, 19) or insect-baculovirus (20) cells against soluble TCRs or by staining T cells with libraries of pMHC-tetramers (21, 22). These approaches are highly valuable, but each has caveats, including time-consuming bacterial purification/refolding of soluble TCRs or expensive synthesis of MHC tetramers. Finally, these methods do not test the most important aspect of T-cell therapy: killing of a target cell.

Here, we have developed a mammalian minigene-based method (termed “PresentER”) of encoding libraries of MHC-I peptides. A PresentER minigene encodes a single peptide that is translated directly into the endoplasmic reticulum using a signal sequence. PresentER-encoded peptides bypass the endogenous protein processing steps that produce MHC-I peptides from full-length proteins. For the purpose of identifying the targets of T cells, the approach described herein is superior to heterologous expression of full-length cDNA as it avoids the unpredictable effects of peptide processing including proteasome cleavage, transporter associated with antigen processing (TAP), and N-terminal trimming by aminopeptidases. PresentER-encoded peptides are noncovalently bound to MHC, as occurs in real cells, as opposed to using a flexible linker covalently tethered to an engineered MHC molecule.

PresentER encoded MHC-I peptides could be used in biochemical and live-cell cytotoxicity assays to sensitively and specifically identify the on-target and off-target ligands of soluble TCR multimers, TCRm antibodies, and engineered TCRs expressed on lymphocytes. PresentER-expressing cells could also be used in immunologic assays such as T-cell activation, cytotoxicity, and tumor rejection in naïve, wild-type mice (23). Using PresentER in pooled library screens, we were able to rapidly discover dozens of peptide ligands of two soluble TCRm antibodies, and two engineered TCR T cells, from among thousands of potential epitopes. For each of the 4 TCR agents studied in this manuscript, multiple cross-reactive peptides were identified, including some which were presented on human cells.

The T2, HEK293T Phoenix-AMPHO, and JY B lymphoblastoid cell lines were purchased from ATCC (CRL-1992; ATCC CRL-3213; ATCC 77441) in 2016. The RMA/S cell line was a generous gift from Dr. Andrea Schietinger (Immunology Program, Memorial Sloan Kettering Cancer Center, New York, NY) in 2016. The TPC1 cell line was a gift from Dr. James Fagin (Human Oncology & Pathogenesis Program, Memorial Sloan Kettering Cancer Center, New York, NY) in 2016. GP2-293 packaging cell lines were purchased from ClonTech (catalog no. 631458) in 2017. These cell lines were not reauthenticated. T2 cells were cultured in Iscove's modified Dulbecco medium supplemented with 10% FBS. RMA/S were cultured in RPMI1640 supplemented with 10% FBS. HEK293T cells were cultured in DMEM supplemented with 10% FBS and 2 mmol/L l-glutamine. TPC1 cells were cultured in DMEM supplemented with 5% FBS and 2 mmol/L l-glutamine. Cultured cells were regularly tested for Mycoplasma. C57BL6/N mice were purchased from Envigo (Envigo:44) and used at 6 to 24 weeks old. OT-1 mice were purchased from Jackson laboratories [C57BL/6-Tg (TcraTcrb)1100Mjb/J; Jackson 003831] and used at 6 to 24 weeks old. Experiments involving animals were conducted with the approval of an Institutional Animal Care and Use Committee at Memorial Sloan Kettering Cancer Center (protocol #96-11-044).

The mouse stem cell virus vector named “MLP” was a generous gift from Dr. Scott Lowe (Cancer Biology & Genetics Program, Memorial Sloan Kettering Cancer Center, New York, NY). The 98 amino acid ENV_MMTVC signal peptide from Mouse Mammary Tumor Virus envelope protein (accession #Q85646; Supplementary Table S1) was found in the Signal Peptide Database (http://www.signalpeptide.de/index.php). A DNA construct referred to as the PresentER Cassette was designed as follows: DNA encoding the ENV_MMTVC signal peptide was codon optimized using the Integrated DNA Technologies (IDT) Codon Optimization Tool and an SfiI restriction site was addednear the 3′-end of the DNA encoding the signal peptide. In doing so, two amino acids were modified: …PQTSLTLFLALL[S>A]VL[G>A]PPPVSG (Supplementary Table S1). A 223-nt filler sequence was encoded downstream of this sequence and followed by a second SfiI site. Double-stranded DNA encoding the modified signal peptide and filler sequence was synthesized by IDT. The DNA was digested with XhoI (NEB, catalog no. R0146) and EcoRI (NEB, catalog no.R3101) and ligated into the XhoI/EcoRI–digested MLP vector using T4 DNA ligase (NEB, catalog no.M0202). The PresentER cassette vector and map are available on Addgene (catalog no.102942).

DNA primers encoding individual PresentER minigene antigens were ordered from IDT in the following format: 5′-GGCCGTATTGGCCCCGCCACCTGTGAGCGGG[…]TAAGGCCAAACAGGCC-3′, amplified with PresentER-F and PresentER-R primers (Supplementary Table S1), digested with SfiI (NEB, catalog no.R0123) and purified with the Qiagen MinElute kit (catalog no. 28004). The PresentER Cassette was digested with SfiI, treated with calf intestinal phosphatase (NEB, catalog no. M0290) and purified by agarose gel electrophoresis. The amplicons were ligated into the digested PresentER backbone with T4 ligase and transformed into NEB stable cells. The PresentER cassette and several example PresentER minigenes are available on Addgene (catalog nos. 102946, 102945, 102944, 102943, 102942, and 102947). Supplementary Table S2 has a list of PresentER constructs used in this article.

Pools of oligonucleotides were synthesized by CustomArray, Inc in the following format: 5′-GGCCGTATTGGCCCCGCCACCTGTGAGCGGG…[24–30 nt insert]…TAAGGCCAAACAGGCC-3′. The oligonucleotides were cloned into the PresentER vector in the same way as the individual minigene. After ligation, the ligation products were electroporated into ElectroMAX DH10B cells (Invitrogen, catalog no. 18290015) following the manufacturer's instructions. Electroporated cells were plated onto four 15 cm² ampicillin plates. Serial dilutions of electroporated cells were plated onto 10 cm² plates to determine the transformation efficiency and number of unique transformed cells. After overnight growth, the colonies were scraped off the plate and grown for 3.5 hours in Terrific Broth (Sigma-Aldrich, catalog no. T0198) supplemented with 1 μg/mL ampicillin at 37°C at 225 rpm and plasmid DNA was prepared with the Qiagen MaxiPrep Kit (Qiagen, catalog no. 12162). Library representation was checked by Illumina sequencing.

HEK293T Phoenix amphoteric cells were transfected with polyethylenimine (PEI) and the PresentER plasmid (15 μg DNA: 45 μg PEI) in 10 cm² tissue culture–treated plates. Virus was harvested every 12 hours, pooled, concentrated with ClonTech Retro-X (Takara, catalog no. 631456). frozen on dry ice, and stored at −80°C. T2 cells were spinoculated with virus in non-TC–treated 6-well plates at 32°C × 2,000 × g for 2 hours in complete media supplemented with 4 μg/mL of polybrene. RMA-S cells were spinoculated with virus in non-TC–treated 6-well plates at 32°C × 1,000 × g for 2 hours in complete media supplemented with 4 μg/mL of polybrene. Cell media were completely exchanged 12 to 24 hours after transduction. Library retrovirus was produced in the same way, except that virus production was scaled up to four 15 cm² plates. The volume of viral supernatant that led to one-third maximal transduction efficiency was established for each batch of virus produced by transduction of target cells with serial dilutions of viral supernatant. Transduced cells were selected with 1 μg/mL (T2) or 4 μg/mL (RMA/S) of puromycin for 2 to 3 days.

The peptides included in the PresentER library were found in Uniprot TrEMBL database of reviewed and unreviewed human protein sequences. Substrings of unique 9 and 10 amino acid sequences were collected and affinity to HLA-A*02:01 was calculated using NetMHCPan. All human peptides with predicted HLA-A*02:01 IC₅₀ less than 500 nmol/L were compared with the ESK1 and Pr20 cross-reactivity motifs to determine which should be included in the library. All potentially cross-reactive ESK1 ligands and half of the potential Pr20 ligands were included in the library.

Purified ESK1 and Pr20 mAbs were provided by Eureka Therapeutics and fluorescently labeled using Innova Biosciences lightning link kits (705-0010) following the manufacturer's instructions. The TCR multimer specific to NLVPMVATV (CMV aa 495–503)/HLA-A*02:01 was purchased from Altor BioScience (catalog no. TCR-CR1-0020). APC-labeled antibodies specific to SIINFEKL/H2-Kb (clone 25-D1.16) were purchased from eBioscience (catalog no. 141606). Each labeled antibody or TCR multimer was titered by flow cytometry on antigen-positive and antigen-negative cells. Briefly, cells were stained with decreasing concentrations of antibody and the concentration with optimal signal:background ratio was identified. Cells were stained as follows: cells were washed twice with ice-cold PBS and then incubated for 30 minutes on ice with fluorescently labeled antibody in staining buffer (PBS supplemented with 2% FCS, 0.1% sodium azide, and 5 mmol/L EDTA). Following staining, cells were washed twice with ice-cold staining buffer and resuspended in staining buffer with DAPI. Fluorescence data was collected on LSR Fortessa (BD Biosciences) or Accuri C6 (BD Biosciences) instruments. Data were analyzed on FlowJo v10.

Fluorescence-activated cell sorting (FACS) was performed to isolate ESK1 and Pr20 bound T2 cells. T2 cells transduced with library virus were stained with ESK1-APC or Pr20-APC as described above. ESK1 or Pr20-positive and -negative cells were sorted on a FACSAria (BD Biosciences) instrument. Sorting gates were set-up based on the fluorescence intensity of stained single-minigene control cells (RMF and ALY). Each sort was performed two times for each antibody. Sorted cells were collected and genomic DNA was purified as described below.

Genomic DNA was extracted from bulk or sorted cells with the Gentra Puregene kit (Qiagen #158689) according to the manufacturer's instructions. Genomic DNA was amplified with Q5 Polymerase (NEB, catalog no. M0491) directly using P5Primer and P7BarcodePrimer or with nested PCR (outer primers P5 and MMTV_SP2_R; inner primers PresentER_Short_F and PresentER_Short_R; Supplementary Table S1). Amplicons were quantitated with PicoGreen and quality control performed by Agilent TapeStation. Sequencing libraries were prepared from amplicons using the KAPA HTP Library Preparation Kit (Kapa Biosystems, KK8234) according to the manufacturer's instructions with 8 cycles of PCR. Barcoded libraries were pooled equimolarly and run on a HiSeq 4000 in a 50 bp/50 bp paired end run, using the HiSeq 3000/4000 SBS Kit (Illumina).

Spot-synthesized crude peptides (PepTrack libraries) were ordered from JPT Peptide Technologies for validation of screen hits. Peptides were resuspended in DMSO to 20 mg/mL and then diluted to 1 mg/mL with PBS. Peptides were pulsed onto T2 cells by supplementing the media of T2 cells with 20–50 μg/mL of each peptide. Pulsed cells were stained with ESK1 or Pr20, as described above, and binding was evaluated by flow cytometry.

JY and TPC1 cells were washed twice with ice-cold PBS, blocked with 1:10 human FcR Block (Miltenyi Biotec, catalog no. 130-059-901) and stained with unlabeled ESK1 or IgG1 isotype control in staining buffer (Eureka Therapeutics, catalog no. ET901) for 30 minutes on ice. Cells were washed with staining buffer and labeled with anti-human IgG1 APC antibody (BioLegend, catalog no. HP6017). Cells were washed twice more in staining buffer and evaluated by flow cytometry.

The A6 and B7 TCRs (Supplementary Table S3) were a generous gift from Dr. Brian Baker (Department of Chemistry and Biochemistry, University of Notre Dame, Notre Dame, IN). The A6 beta-chain plasmid was modified to encode a C-terminal AviTag biotinylation site (GLNDIFEAQKIEWHE). Gibson cloning was used to insert the AviTag site with the following two primers: F: 5′-gcagaaaattgaatggcatgaaTAAGCTTGAATTCCGATCCGG-3′ R: 5′-gcttcaaaaatatcgttcaggccGTCTGCTCTACCCCAGGC-3′. Both the alpha and beta chains were expressed in BL21 (DE3) bacteria (Thermo Fisher Scientific, catalog no. C606010) and induced with 1 mmol/L IPTG (Sigma-Aldrich, catalog no. 16758). The A6 beta chain was cotransfected with a plasmid encoding the BirA enzyme (Addgene, catalog no. 26624) and bacteria were grown with supplemental biotin (0.5 mmol/L, D-Biotin). Inclusion bodies were harvested and the individual chains were purified and refolded together according to the previously described protocols (24, 25).

To generate full-length mammalian TCR sequences, we used Gibson assembly (NEB, catalog no. E2611) to clone the A6 and B7 alpha/beta chains (Supplementary Table S3) into the pMSGV1 vector backbone.

Written consent to Institutional review board protocol #06-107, conducted according to the common rule, was obtained from four donors over the age of 18 and meeting the inclusion criteria of “healthy participant with no current malignancies.” Blood was collected by venipuncture into heparinized tubes. Peripheral blood mononuclear cells (PBMC) were isolated by differential density centrifugation using Lymphocyte Separation Medium (Corning, catalog no. 25072CI) within 24 hours.

Plasmids encoding the DMF5 and 1G4 TCRs (Supplementary Table S4) were provided as a kind gift from Dr. Steven A. Rosenberg (Surgery Branch, National Cancer Institute, Center for Cancer Research, National Institutes of Health, Bethesda, MD). TCR retroviral transduction was performed as described previously (26). Briefly, retroviral particles were generated by transient transfection of the retroviral packaging cell line 293GP cells with the pMSGV1-TCR plasmids and pRD114 plasmid using Lipofectamine 2000 (Life Technologies). Retroviral supernatant was harvested 2 days later and used to transduce PBMCs that were stimulated with soluble 50 ng/mL anti-CD3 (clone OKT3, Miltenyi Biotec, catalog no. 130-093-387) and 300 IU/mL rhIL2 (R&D Systems) for 2 days prior to retroviral transduction. Retroviral transductions were performed on Retronectin (Takara) coated nontissue culture–treated 24-well plates by spinoculation of the retrovirus at 2,000 × g, 32°C for 2 hours, followed by addition of activated T cells to the retrovirus-containing plates. After overnight incubation at 37°C, T cells were transferred to a tissue-culture treated 24-well plate and expanded in human T-cell media (RPMI supplemented with 10% FBS, 1% l-glutamine, 300 IU/mL rhIL2). Transduced T cells were used at 10–15 days posttransduction or cryopreserved until used in assays.

IFNγ release ELISpot assays were performed in 200 μL of RPMI supplemented with 5% FBS. Fifty-thousand F5-transduced T cells were incubated at 1:1 effector:target (E:T) ratios with T2 cells expressing PresentER minigenes. Controls included no targets, wild-type T2s, 50 μg/mL peptide-pulsed T2s, and PHA-treated cells. Coculture assays were performed in a similar manner: 50,000 T cells were cocultured in U-bottom plates with mixtures of 50,000 T2 cells expressing PresentER minigenes in mCherry or GFP flavors. Control wells did not contain T cells. Antigen-specific target cell lysis was evaluated by flow cytometry 45 hours after coculture by comparing the percentage of GFP-positive cells to the percentage of mCherry cells in each well.

Mouse RMA/S cells were transduced at low multiplicity of infection (MOI; <0.3) with a library of 5,000 PresentER minigenes encoding wild-type H-2Kb peptides (NetMHCPan v4.0 predicted IC₅₀ < 500 nmol/L) selected randomly from the mouse proteome (UniProt database UP000000589 of canonical mouse protein sequences) together with several control minigenes encoding known H-2Kb antigens (e.g., chicken ovalbumin SIINFEKL). Transduced cells were selected with puromycin as described above. Spleens from OT-1 or C57B6 mice were removed aseptically, grinded against a 100-μm filter, washed with PBS, and then cultured in mouse T-cell media (RPMI supplemented with 10% FBS + 1% HEPES + 1% sodium pyruvate + 1% HEPES + 1% l-glutamine + 50 μmol/L 2-mercaptoethanol + 200 IU/mL hlL2). Splenocytes were activated by adding 1 μg/mL anti-CD3e (clone 145-2C11; BD Biosciences, catalog no. 553058) and 1 μg/mL anti-CD28 (clone 37.51; BD Biosciences, catalog no. 553295) to the media. RMA/S-expressing library minigenes were cocultured with activated OT-1 or activated C57B6 splenocytes for 4 days in 96-well round bottom plates. To prevent media exhaustion and avoid well-specific effects, all wells were pooled together daily, fresh media were added, and cells were redispersed onto new 96-well plates. After 96 hours, the cells were removed from the plate, genomic DNA was extracted, and minigenes were quantified by Illumina sequencing as described above.

The human T2 library depletion assays were performed in a similar fashion. T2 cells were transduced with minigene libraries encoding HLA-A*02.1 peptides. Library expressing cells were cocultured with A6, B7, DMF5, or 1G4 T cells for 4 days. Cells were pooled together daily, media were added as needed, and cells were redispersed onto the plates to prevent media exhaustion and avoid well-specific effects. Finally, cells were harvested, genomic DNA was extracted, and minigenes were quantified by Illumina sequencing as above.

Affinity columns were prepared as follows: 40 mg of cyanogen bromide–activated-Sepharose 4B (Sigma-Aldrich, catalog no. C9142) was activated with 1 mmol/L hydrochloric acid (Sigma-Aldrich, catalog no. 320331) for 30 minutes. Subsequently, 0.5 mg of W6/32 antibody (BioXCell, BE0079; RRID: AB_1107730) was coupled to Sepharose in the presence of binding buffer (150 mmol/L sodium chloride, 50 mmol/L sodium bicarbonate, pH 8.3; sodium chloride: Sigma-Aldrich, catalog no. S9888, sodium bicarbonate: Sigma-Aldrich, catalog no. S6014) for at least 2 hours at room temperature. Sepharose was blocked for 1 hour with glycine (Sigma-Aldrich, catalog no. 410225). Columns were equilibrated with PBS for 10 minutes.

T2 cells expressing PresentER constructs were washed three times in ice-cold sterile PBS. Afterwards, cells were lysed in 7.5 mL 1% CHAPS (Sigma-Aldrich, catalog no. C3023) in PBS, supplemented with protease inhibitors (Roche cOmplete, catalog no. 11836145001) for 1 hour at 4°C. Cell lysate was centrifuged at 20,000 × g for 1 hour at 4°C. Supernatant was passed over column through peristaltic pumps at 1 mL/minute flow rate overnight at 4°C. Affinity columns were washed with PBS for 30 minutes, water for 30 minutes, then run dry, and HLA complexes subsequently eluted five times with 200 μL 1% trifluoracetic acid (TFA, Sigma/Aldrich, catalog no. 02031). For separation of HLA ligands from their HLA complexes, tC18 columns (Sep-Pak tC18 1 cc Vac Cartridge, 50 mg Sorbent per Cartridge, 37–55 μm Particle Size, Waters, catalog no. WAT036820) were prewashed with 80% acetonitrile (ACN, Sigma-Aldrich, catalog no. 34998) in 0.1% TFA and equilibrated with two washes of 0.1% TFA. Samples were loaded, washed again with 0.1% TFA, and eluted in 400 μL 30% ACN in 0.1% TFA. Sample volume was reduced by vacuum centrifugation for mass spectrometry analysis.

Samples were analyzed by high resolution/high accuracy LC/MS-MS (Lumos Fusion, Thermo Fisher Scientific). Peptides were desalted and concentrated prior to being separated using direct loading onto a packed-in-emitter C18 column (75 μm ID/12 cm, 3-μm particles, Nikkyo Technos Co., Ltd.). The gradient was delivered at 300 nL/minute increasing linear from 2% Buffer B (0.1% formic acid in 80% acetonitrile)/98% Buffer A (0.1% formic acid) to 30% Buffer B/70% Buffer A, over 70 minutes. MS and MS/MS were operated at resolutions of 60,000 and 30,000, respectively. Only charge states 1, 2, and 3 were allowed. 1.6 Th was chosen as isolation window and collision energy was set at 30%. For MS/MS, maximum injection time was 100 ms with an AGC of 50,000.

MS data were processed using Byonic software (version 2.7.84, Protein Metrics). Mass accuracy for MS1 was set to 6 ppm and to 20 ppm for MS2, respectively. Digestion specificity was defined as unspecific and only precursors with charges 1, 2, and 3 and up to 2 kDa were allowed. Protein FDR was disabled to allow complete assessment of potential peptide identifications. Oxidization of methionine was set as variable modifications for all samples. All samples were searched against UniProt Human Reviewed Database (20,349 entries, http://www.uniprot.org, downloaded June 2017). Peptides were selected with a minimal log probability value of 2 resulting in a 1% false discovery rate. For visualization of mass spectrometry results skyline software (version 3.1, MacCoss Lab Software) was used. Masses of precursors and product ions of peptide target sequences were searched in all relevant .raw files and peak areas of all replicates compared.

After Illumina sequencing of each sample, reads were mapped to the appropriate PresentER minigene library with Bowtie2. Reads that did not map to the minigenes in the library were discarded. For each sample, the number of reads mapping to each minigene was divided by the total number of reads that mapped to the library. Minigenes with frequencies less than 1/50,000 to 100,000 reads in untreated or unsorted samples were excluded from further analyses. To identify minigenes enriched in the screen, we normalized the frequency of each minigene in the sorted/treated samples by the frequency of each minigene in the unsorted/untreated samples. For experiments with T-cell cocultures, the frequency of each minigene was compared between treatment groups (e.g., A6 cocultured libraries vs. 1G4 cocultured libraries). The A6 and B7 enrichment/depletion scores were calculated by dividing the frequency of minigenes in the library after coculture with A6 or B7 divided by the frequency of minigenes in the library prior to coculture. For the ESK1/Pr20 enrichment/depletion scores, the frequency of each minigene in the sorted “antibody high” sample was divided by the minigene frequency in the “antibody low” sample.

Statistical analysis was performed in Prism v7 and in the R programming language. Statistical tests are described in the figure legends and include two-tailed t tests with a significance cutoff of P ≤ 0.05. Receiver operating curves were generated using the R package “plotROC” (Sachs:2017hx).

We designed a minigene (“PresentER”) that is capable of generating precisely defined MHC-I antigens in mammalian cells. DNA encoding the MHC-I peptide is short (72–78 nt) and inexpensive to synthesize individually or as a pool (Fig. 1A; Supplementary Fig. S1). The peptide is encoded downstream of a signal sequence, thus bypassing the typical processing for MHC-I peptide presentation: proteasomal cleavage, peptide transport into the endoplasmic reticulum, and aminopeptidase trimming (27). The peptide is translated directly into the endoplasmic reticulum, where it binds to MHC and is exported to the surface of the cell. PresentER was designed for pooled screening applications; therefore, the amino acid sequence corresponding to an MHC-presented peptide is encoded by DNA only once per minigene and can be sequenced in its entirety in one 50 bp sequencing read. To demonstrate that PresentER minigenes recapitulated all of the expected characteristics of MHC-I–presented antigens, we relied on several fluorescently labeled mAbs, multimerized TCRs, and engineered T cells. All TCR and TCR-like agents used as reporters of cell surface MHC-I are described in Table 1.

We have previously isolated and characterized two TCRm antibodies that recognize HLA-A*02:01 presented peptides derived from genes aberrantly expressed on cancer cells. ESK1 (5) binds to the Wilms tumor (WT1) peptide RMFPNAPYL (WT1:126-134) and Pr20 (28) binds to the preferentially expressed antigen of melanoma (PRAME) peptide ALYVDSLFFL (PRAME:300-309). T2 cell lines expressing PresentER-encoded peptides were bound by ESK1 and Pr20 only when expressing their cognate antigens, but not irrelevant peptides (Fig. 1B; Supplementary Fig. S2A–S2D). To show that TCR could bind to the minigene-derived MHC-I peptides, we stained PresentER antigen–expressing T2 cells with HLA-A*02:01 peptide specific TCR multimers: (i) TCR-CR1 which recognizes NLVPMVATV (cytomegalovirus pp65:495–503) and (ii) A6 which recognizes LLFGYPVYV (HTLV-1 Tax:11–19; ref. 29). These TCR multimers specifically bound T2 cells expressing their cognate ligand (Supplementary Fig. S2E–S2H).

PresentER-encoded minigenes could also lead to presentation of peptides on nonhuman MHC molecules: an antibody against mouse H-2Kb/SIINFEKL bound the Tap2-deficient mouse RMA/S (30) cells expressing PresentER SIINFEKL (chicken ovalbumin 257–264) but not cells expressing PresentER MSIIFFLPL (PEDF:271–279; Fig. 1C). Cell lines with wild-type Tap function could present PresentER-driven antigens, as shown by the expression of PresentER-SIINFEKL in EL4 cells (Supplementary Fig. S3).

We immunoprecipitated peptide–MHC complexes from T2 cells expressing PresentER-RMFPNAPYL or PresentER-ALYVDSLFFL and identified bound peptides by mass spectrometry. RMFPNAPYL and ALYVDSLFFL were identified only in cells encoding those PresentER constructs (Supplementary Fig. S4A and S4B). No peptides derived from the ER signal sequence were identified.

To confirm that signal sequence–mediated delivery of the antigen to the endoplasmic reticulum was the mechanism of MHC-I peptide presentation, we generated minigene constructs with scrambled ER signal sequences. The scrambled signal sequences will not lead to peptide translocation to the ER and thus should abrogate TCRm binding. These minigenes were generated by shuffling the amino acids of the (N-terminal) signal sequence, but not the (C-terminal) peptide antigen sequence. In T2 cells expressing two different scrambled signal sequences coupled to peptide antigens ALYVDSLFFL or RMFPNAPYL, we found no binding of ESK1 or Pr20 (Supplementary Fig. S4C).

Pooled screens using retroviral methods must be “single-copy competent,” meaning that a single DNA copy of the minigene must suffice to yield the desired effect in a cell. If more than one copy is required, the pooled screen will fail because few cells will receive more than one copy of the same minigene (31). To demonstrate that PresentER minigenes were single copy-competent, we infected mouse H-2Kb RMA/S cells with decreasing titers of virus carrying PresentER SIINFEKL (i.e., decreasing multiplicities of infection) and then assayed for SIINFEKL/H-2Kb presentation on the surface of the cell using the 25-D1.16 antibody. SIINFEKL peptides were still found on the surface of the cell after 1,000-fold viral dilution and resultant 15-fold decrease in infection rate (<1% of cells infected). This demonstrated that the PresentER minigene was capable of driving antigen presentation from a single copy of the retroviral minigene, thereby enabling its use in a pooled screen (Supplementary Fig. S5).

We next tested whether PresentER-encoded MHC-I antigens could be used in functional T-cell assays. In addition to the A6 TCR described above, other genetically engineered T cells to known antigens are described in the literature, including DMF5 (refs. 3, 32; specific to MART-1:27-35 HLA-A*02:01/AAGIGILTV), 1G4 (ref. 33; specific to NY-ESO-1:157-165 HLA-A*02:01/SLLMWITQC), and B7 (same specificity as the A6 TCR). We transduced T cells from non-HLA-A*02:01 donors with constructs encoding the DMF5, 1G4, or A6 TCRs and then cocultured these cells with peptide-pulsed or PresentER minigene–expressing T2 cells. The T cells released similar amounts of IFNγ when exposed to minigene or peptide-pulsed cells, demonstrating that PresentER minigenes yield pMHC at levels sufficient to lead to T-cell recognition (Fig. 1D).

To confirm that engineered T cells could specifically kill T2 cells expressing their cognate antigen, we cloned several PresentER minigenes into a PresentER vector encoding mCherry instead of GFP. Then, we mixed T2 cells expressing peptide A/GFP with T2 cells expressing peptide B/mCherry. For instance, T2 cells expressing PresentER Tax (GFP) were mixed with T2 cells expressing PresentER MART-1 (mCherry). These mixtures were then cocultured with T cells specific to one of the antigens, or with T cells specific to an irrelevant antigen. After 45 hours, flow cytometry was used to evaluate the percentage of live cells expressing each fluorophore. T cells killed T2 cells expressing their cognate antigens, but not T2 cells expressing an irrelevant antigen (Fig. 1E). This experiment confirmed that T cells selectively eliminated T2 cells expressing their PresentER-encoded cognate antigen.

To test whether PresentER libraries could be used to distinguish MHC-I targets of T cells from irrelevant targets at large scale, we used T cells from OT-1 mice, which express only H-2Kb/SIINFEKL–specific TCRs, and T cells from B6 mice, which are poly-specific. Activated splenocytes from OT-1 and B6 mice were cocultured with RMA/S cells expressing a library of 5,000 randomly selected wild-type H-2Kb peptides as well as several control peptides, including SIINFEKL. The abundance of each minigene after coculture was assayed by Illumina sequencing. A schematic of the experiment is presented in Fig. 2A. Cells expressing the peptide target of the OT-1 TCR—SIINFEKL—were depleted by more than 1-log while no other minigenes were depleted (Fig. 2B).

Next, we tested the ability of the PresentER method to discover the targets and off targets of two human TCRs. The A6 and B7 TCR were isolated from a patient with HTLV-1-associated myelopathy/tropical spastic paraparesis (HAM/TSP) and were found to recognize an HLA-A*02:01 peptide derived from the HTLV-1 virus Tax protein. The A6 TCR, in complex with its target, was the first human TCR structure to be solved and the biochemical characteristics of both A6 and B7 have been extensively explored (24, 25, 34). The position-specific binding specificities for the A6 and B7 TCR are mapped by cytotoxicity studies using peptide-pulsed target cells (35). These features made the A6 and B7 TCRs excellent candidates to use in a screen to validate previously known targets and discover additional targets. Using published data, (35) we generated a position-specific scoring matrix (PSSM) to predict A6 or B7 targets. We scored the human proteome according to the PSSM to find peptides that might be targets of A6 and B7. We selected 5,000 peptides with high predicted affinity to HLA-A*02.1 that were either highly scored for A6, B7, or both. We also included all single amino acid mutants of LLFGYPVYV. We cloned a library of minigenes encoding these peptides and introduced them into T2 cells. T cells from non-HLA-A*02.1 donors were transduced with plasmids encoding the A6, B7, DMF5, or 1G4 TCR and cocultured with the library of target cells. The abundance of each minigene was quantified by Illumina sequencing before and after coculture with T cells expressing each TCR. Fifty-two minigenes were depleted by 2-fold or more after coculture with A6 T cells and 34 minigenes were similarly depleted by coculture with B7 T cells. No minigenes were depleted by DMF5 or 1G4 (Fig. 2C). To demonstrate the interexperimental reliability of minigene depletion by T-cell coculture, minigene depletion by A6 T cells from two different PBMC donors was compared. The minigenes that were depleted more than 2-fold between the two donors were 90% concordant (Fig. 2D). All the depleted minigenes were single amino acid substitutions of the Tax peptide. As expected, based on prior data (35), the A6 TCR was more promiscuous, recognizing almost twice as many peptides as B7 (Fig. 2E).

To examine the ability of the PresentER screen to discover peptide targets of A6 and B7, we focused on the minigenes encoding single amino acid variants of the Tax peptide. Many of these peptides were previously tested for A6 and B7 cytotoxicity by Hausmann and colleagues (35). A PSSM depicting this is presented in Fig. 3A and B. Many A6 and B7 ligands validated by Hausmann and colleagues (35) were depleted in our coculture assays (Fig. 3C and D). We identified substitutions that were not tested by Hausmann and colleagues (35) but were found to be targets of the TCRs; 19 substitutions led to A6 cytotoxicity and 17 substitutions led to B7 toxicity. These included substitutions at the 2nd and 9th position (revealing unusual MHC anchors), substitutions containing cysteines and differences between glutamic acid, and aspartic acid residues in the same positions (Fig. 3E and F).

To assess the sensitivity and specificity of a TCR target screening tool such as PresentER, several targets and nontargets must be known for each TCR. For most TCRs, few peptide ligands are known because testing TCR reactivity for a meaningfully large set of synthetic peptides is cost prohibitive. The A6 and B7 TCR are unique in that a large number of peptides have already been tested: 105 out of 133 possible single amino acid mutants of the LLFGYPVYV peptide were tested by Hausmann and colleagues (35) in coculture killing assays. Using Hausmann's data as the “ground truth,” we categorized peptides with >25% of maximum killing as “targets” and the rest as “nontargets” and plotted ROC of the PresentER A6 and B7 screens (Fig. 3G). The A6 TCR screen was highly sensitive, reaching >90% sensitivity with 85% specificity. The B7 TCR screen was less effective, reaching 70% sensitivity with 85% specificity. 1G4 is plotted for comparison. The sensitivity of PresentER was likely underestimated by this analysis, as the nonphysiologic conditions of peptide pulsing in vitro tends to overestimate the T-cell activating potential of a peptide.

Two peptides that are characterized as weak off targets of the A6 and B7 TCRs in peptide-pulsing assays were also included in the library: S. Cerevisiae Tel1P 549–557 MLWGYLQYV and Human HuD/ELAVL4 87–95 LGYGFVNYI (35). Neither of these peptides were depleted in the minigene library depletion assays, prompting us to wonder whether these peptides were indeed recognized well by the A6/B7 T cells. We performed an ELISpot using T2 cells pulsed with peptide or expressing PresentER minigenes. A6 was weakly reactive to the ELAVL4 peptide when pulsing onto T2 cells, but not to ELAVL4 minigene–expressing cells. A6 did not react to Tel1p peptide pulsed cells or Tel1p minigene–expressing cells. B7 was not reactive to Tel1p or ELAVL4 peptides or PresentER minigenes by ELISpot (Supplementary Fig. S6A). T2 coculture killing assays using minigene-expressing peptides were also negative (Supplementary Fig. S6B). These results suggested that in some cases there may be differences between the peptides that could be pulsed onto cells and peptides that could be presented via a minigene.

Two anticancer antibodies developed by our group, Pr20 and ESK1, have been extensively evaluated in preclinical studies as therapeutic agents. These antibodies were developed using phage display libraries and thus never underwent a thymic-like negative selection process. Therefore, it is unknown whether they might cross-react with peptides presented endogenously on human cells. On the basis of alanine/residue scanning and structural (36) data, ESK1 binding to RMFPNAPYL depends primarily on the R1 and P4 residues. In contrast, Pr20 binds to the C-terminus of the peptide (28). Therefore, we constructed a library of possible ESK1 and Pr20 cross-reactive targets by searching the human proteome in silico for 9 and 10-mer peptide sequences matching a motif based on prior biochemical data (Fig. 4A). We located 1,157 and 24,539 potential cross-reactive peptides of ESK1 and Pr20, respectively, with NetMHCPan (37)-predicted HLA-A*02:01 affinity of less than 500 nmol/L (Fig. 4B). We synthesized a library of 12,472 oligonucleotides that together encoded all of the ESK1 cross-reactive peptides, half of the Pr20 cross-reactive targets, the single amino acid mutants of RMF and ALY (termed “CR-ESK1” and “CR-Pr20”, respectively), and positive/negative control peptides (Fig. 4B). Library-transduced T2 cells were stained with ESK1 or Pr20, sorted, and sequenced. Every minigene was scored for ESK1 and Pr20 enrichment, indicating how frequently it was found in the highly stained cells compared with the poorly stained cells. A schematic of the flow-based screen is presented in Fig. 4C.

First, we wanted to determine whether previously known ESK1 ligands were enriched in the sorted cells. Minigenes encoding known ESK1 ligands had higher enrichment scores compared with nonligands (P = 0.032). This suggested that the flow-based screen was able to separate ESK1 binders from nonbinders (Supplementary Fig. S7A). Next, we looked to see which minigenes were enriched in the screen. We found over 100 peptides with IC₅₀ < 500 nmol/L which were >5-fold enriched for ESK1 binding. Surprisingly, several of the most enriched peptides that emerged in the ESK1 screen were CR-Pr20 peptides, such as RVIMPCNWWV and RMFSGVGVYL (Fig. 4D). Although these peptides are 10-mers, some bear sequence similarity to the target ligand of ESK1. To validate these hits, we synthesized 27 of the enriched peptides, pulsed them onto T2 cells, and stained them with ESK1. Of the peptides tested, 22 (81%) showed binding to ESK1, including several which had originally been selected for Pr20 cross-reactivity and did not contain a proline in position 4 (Fig. 4E). These unusual targets could not have been predicted from either the crystal structure of ESK1 or the alanine scanning data.

Next, we wanted to determine whether any of the ESK1 targets we had discovered were expressed in a WT1-negative cell line, thus possibly leading to ESK1 binding. Large databases of HLA-A*02:01 peptide ligands isolated from tumors and normal tissue have become available (18, 38–40). Within these databases (including personal correspondence with Department of Immunology members at Tübingen), we found two WT1-negative (41) cell lines that contained ESK1 off-targets discovered in the PresentER screen (TPC-1: RLPPPFPGL, RVMPSSFFL, RLGPVPPGL, JY: KLYNPENIYL, RLVPFLVEL). RMFPNAPYL was not found among the MHC-I ligands immunoprecipitated from these lines. We tested ESK1 binding in each of these lines and found that JY cells bound ESK1 at high intensity, whereas TPC-1 was marginally positive for ESK1 binding (Fig. 4F and G). Thus, PresentER may be used to identify both theoretical and, in some cases, peptides presented on real cells that are bound by ESK1.

A screen of Pr20 cross-reactive ligands was performed in the same manner as described for ESK1. Known Pr20 binders were not enriched relative to the negative controls (P = 0.71; Supplementary Fig. S7B). However, 20 peptides were more than 5-fold enriched for Pr20 binding with predicted IC₅₀s of less than 100 nmol/L. We synthesized 13 of these peptides, pulsed them onto T2 cells, and found that all 13 bound to Pr20 (Supplementary Fig. S8).

Examining only the CR-ESK1 subset of peptides, we noticed that the peptides most enriched for ESK1 binding were also predicted to have the highest affinity for HLA-A*02:01 (Fig. 5A). Peptides that are ≥5-fold enriched for ESK1 binding have a higher affinity for HLA-A*02:01 as compared with the library as a whole and compared with the peptides that were ≥5-fold depleted (median affinity of 31 nmol/L, 95 nmol/L and 102 nmol/L, respectively; Fig. 5B). We found the same result in the Pr20 screen: the most enriched Pr20 ligands also had the highest affinity to MHC-I. (Fig. 5C). The skew we observed in both ESK1 and Pr20-enriched minigenes toward high-affinity HLA-A*02:01 ligands suggests that minigene expression of peptides selects for presentation of ligands with the highest affinities for HLA-A*02:01. This may be an unexpected feature of PresentER, as affinity to MHC-I is the most important factor in endogenous MHC-I peptide presentation (although high peptide expression levels may overcome low affinity; ref. 17). We cloned minigenes for four of the most enriched CR-ESK1 (RLFPLAWTV 31.8×; KLMGAISFFI 41.9×) and CR-Pr20 (WLLGDSSFFL 6.5×; LLIQEGPFFV 6.6×) peptides and tested them for binding to ESK1 and Pr20. ESK1- and Pr20-bound cells presenting these peptides at significantly higher amounts than the peptides to which these antibodies were originally isolated (Fig. 5D).

Few methods exist for robust identification of the targets of TILs and the off-targets of TCR-based therapeutic agents and cells. As a consequence, preclinical evaluation of novel therapeutic agents directed toward peptide-MHC are insufficient to prevent harmful off-tumor off-target toxicities, including deaths. A number of approaches can identify the off targets of tumors, but all suffer from key limitations. For instance, animal models of cross-reactivity are not very useful due to species-specific MHC molecules and differences in peptide processing (42). Yeast/Insect display is highly scalable, but relies on purified and refolded TCRs in vitro, which is not the native format that would be delivered to patients. Screening using MHC tetramers can be effective, but synthesis of large numbers of peptides is expensive. Methods to generate DNA-barcoded tetramers using in vitro transcription and translation are promising and we look forward to future development in this area. However, while these methods are valuable and can elucidate the fundamental biology of TCRs (43), they are poorly suited to preclinical evaluation of novel therapeutics.

PresentER is a mammalian cell–based approach to identify MHC-I ligands of TCR agents at large scale. The PresentER method is a physiologically relevant, scalable method that can be used to identify functional cross-reactivities between MHC-I ligands and TCR agents with excellent sensitivity and specificity. In this report, we have demonstrated that it can also be used for immunologic assays in vitro while in a separate publication we demonstrated its use in immunologic assays in vivo (23).

In this article, we used peptide libraries biased towards existing biochemical binding data to study the off targets of TCR-like molecules. However, users of the PresentER minigene method could instead generate unbiased libraries. As an example, a library of minigenes including all possible 9–10 amino acid peptides in the human exome that bind to one allele of MHC-I (<1 × 10⁶) could be constructed for $50,000, compared with the many months to years of work and the millions of dollars needed to do the same with synthesized peptides. This is a trivial sum for the preclinical evaluation of a novel engineered T-cell therapeutic compared with the cost of a clinical trial that is halted because of toxicity (8, 9).

We found that some peptides act as TCR/TCRm ligands when pulsed onto T2 cells at supraphysiologic levels, but are not detected when expressed by PresentER minigenes (e.g., A6 and its target ELAVL4). Whether or not these peptides are presented endogenously in any human tissue is unknown. We speculate that some PresentER-encoded peptides may never reach the cell surface because they are removed by MHC peptide editors (44) or aminopeptidases. Alternatively, some peptides may be inefficiently loaded onto MHC when they are not transported by Tap or may be unstable once they reach the cell surface. These processes are likely to disproportionately affect peptides with low affinity to MHC. Consequently, the restrictions imposed by cellular machinery may generate more reliable PresentER screening results as compared with peptide pulsing assays. Indeed, we found that peptides with low MHC-I affinity are infrequently off-targets of ESK1 and Pr20 in our screens. However, because of the pooled library approach used here, measurements of affinities and stability are not possible, nor have we been able to identify library peptides by MS/MS, likely because of low absolute abundance. For other applications, the inability to present some low-affinity MHC peptides may be undesirable. For instance, some proteins that are highly expressed in human tissues may lead to endogenous peptide presentation despite the low affinity to MHC. Peptides derived from these proteins may not be properly presented when encoded by PresentER minigenes.

Another caveat to our approach is that a large number of target cells are required to ensure library fidelity, at least 1,000 cells per minigene, thus limiting the size of each minigene library to the number of cells that can be manageably cultured. This is solved by splitting large libraries into multiple sublibraries. Finally, we demonstrated the use of PresentER minigenes for ligands of HLA-A*02:01 and mouse H-2Kb, but not other MHC alleles. In principal, PresentER minigenes should work for any MHC allele because while the peptide binding motifs may differ, the mechanisms of loading and presenting peptides are the same.

Because PresentER screening of TCR relies on T-cell coculture and killing, it may fail to identify pMHC that antagonize TCR function (45). Moreover, there is even a risk that antagonist peptides found in the library could skew the results by inhibiting T-cell activation. However, this is unlikely to be a problem unless the numbers of antagonist peptides in the library is so large as to productively inhibit most of the T cells in the assay.

PresentER can be used for biochemical evaluation of therapeutic TCR-based agents such as engineered TCRs and TCRm antibodies and discovered dozens of off targets of each. We believe that the minigene method we have described meets an unmet need in the preclinical evaluation of TCR agents. Given that patients have already died as a result of off-tumor, off-target toxicity, we propose that off target assessment using libraries of MHC minigenes covering the entire human exome be a routine step in preclinical development of TCR and TCR-like agents.

R.S. Gejman has ownership interest in the patent filed on the PresentER technology (U.S. provisional patent application no. 62/395,577). A.Y. Chang is a postdoctoral research fellow at Pfizer. T. Dao is a consultant/advisory board member for Eureka Therapeutics. C.A. Klebanoff is a scientific advisory board member for Achilles Therapeutics, Aleta Biotherapeutics, Bellicum Pharmaceuticals, Obsidian Therapeutics, Roche/Genentech, Klus Pharma, and G1 Therapeutics, and reports receiving a commercial research grant and other commercial research support from Kite/Gilead. D.A. Scheinberg is an advisor for Eureka Therapeutics and Oncopep, reports receiving a commercial research grant from Bristol-Myers Squibb, has ownership interest (including patents) in Sellas and Iovance, and has ownership interest in the patent filed on the PresentER technology. No potential conflicts of interest were disclosed by the other authors.

Conception and design: R.S. Gejman, C.A. Klebanoff, D.A. Scheinberg

Development of methodology: R.S. Gejman, H.F. Jones, C.A. Klebanoff, D.A. Scheinberg

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): R.S. Gejman, H.F. Jones, M.G. Klatt, A.Y. Chang, C.Y. Oh, S.S. Chandran, V. Zakahleva

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): R.S. Gejman, M.G. Klatt, A.Y. Chang, C.Y. Oh, C.A. Klebanoff, D.A. Scheinberg

Writing, review, and/or revision of the manuscript: R.S. Gejman, H.F. Jones, M.G. Klatt, A.Y. Chang, C.Y. Oh, S.S. Chandran, C.A. Klebanoff, D.A. Scheinberg

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): R.S. Gejman, T. Korontsvit, V. Zakahleva, T. Dao, D.A. Scheinberg

Study supervision: R.S. Gejman, C.A. Klebanoff, D.A. Scheinberg

The authors thank Eureka Therapeutics for the ESK1 and Pr20 antibodies; Brian Baker and Lance Hellman for the A6 and B7 TCR plasmids; Yael David for assistance in refolding soluble TCRs; Scott Lowe and Eusebio Manchado Robles for the MSCV backbone and assistance with pooled library cloning; and Steven A. Rosenberg for the DMF5 and 1G4 TCR. The authors thank the Integrated Genomics Operation Core, funded by the NCI Cancer Center Support Grant (CCSG, P30 CA08748), Cycle for Survival, and the Marie-Josée and Henry R. Kravis Center for Molecular Oncology and the Flow Cytometry core facilities at Memorial Sloan Kettering Cancer Center for their assistance. This project was supported by the Parker Institute for Cancer Immunotherapy and the Functional Genomics Initiative at Memorial Sloan Kettering Cancer Center. R.S. Gejman is supported by NCI F30 CA200327 and National Institute of General Medical Sciences (NIGMS) T32GM07739. D.A. Scheinberg is supported by NCI RO1 CA 55349, P01 CA23766, and P30 CA008748. M.G. Klatt is supported by Deutsche Forschungsgemeinschaft grant no. KL 3118/1-1. C.A. Klebanoff is the recipient of a Damon Runyon Clinical Investigator Award.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.