Introduction

The adoptive transfer of tumor-infiltrating T lymphocytes (TIL) recognizing autologous cancer has yielded remarkable success in metastatic melanoma.1 Studies on clinical responders revealed that TIL often targeted proteins encoded by genes harboring tumor-specific mutations, known as cancer neoantigens.2 3 This led to the development of a clinical protocol involving the isolation, expansion, and administration of neoantigen-reactive TIL to patients with other cancer types.4 5 Subsequently, durable remissions were attained in specific cases of metastatic bile duct,6 colon,7 and breast cancer.8 An alternative strategy involving transduction of T-cell receptors (TCRs) from neoantigen-reactive TIL into autologous circulating T cells also demonstrated therapeutic promise.9 10

Because most cancer neoantigens are unique, selecting TIL or TCRs for adoptive cell therapy (ACT) involves personalized, laborious, and time-intensive screening.11 This process starts with the resection of metastatic tumor(s), which are used for both generating TIL cultures and identifying tumor-specific mutations through next-generation sequencing (NGS). TIL are subsequently screened for recognition of peptides and minigenes representing these mutations, and those recognizing mutant peptides are then expanded for therapy or used to isolate TCR sequences.

A significant limitation of this approach is its reliance on invasive tumor sampling. Furthermore, the synthesis of peptides and minigene screening libraries typically begins only after tumor resection and sequencing, considerably increasing the overall protocol lead time.

Prior studies have explored non-invasive methods to isolate neoantigen-reactive T cells, including screening circulating programmed cell death protein-1 (PD-1)-positive or memory T cells from patients with metastatic cancer.12–14 However, complementary methods that would allow for comprehensive mutational analysis without invasive sampling of active tumors remain unexplored. These include NGS on circulating tumor cells, plasma cell-free DNA, or archived, formalin-fixed, paraffin-embedded (FFPE) tumor samples.

Archived FFPE samples may be readily accessible for patients with metastatic cancer who underwent prior biopsies or resection of their primary tumors. Based on high concordance in cancer driver mutations between matched primary and metastatic tumors, sequencing primary tumors is expected to capture most neoantigens encoded by these mutations.15–17 However, it remains unclear if this applies to neoantigens encoded by passenger mutations, which represent most cancer neoantigens.4 5 Previous work addressing this question focused on mutations predicted to generate class I major histocompatibility complex (MHC)-binding peptides but did not validate them as neoantigens in T-cell assays.18

The primary objective of our study was to investigate the distribution of validated cancer neoantigens between matched primary and metastatic tumors. By doing so, we aimed to assess the feasibility of using FFPE primary tumor samples for cancer neoantigen discovery. This could reduce the need for additional tumor resections (if coupled with non-invasive T-cell isolation methods), expedite screening library preparation, and potentially improve targeted neoantigen selection. Additionally, we wanted to determine the extent to which prior tumor samples shared the mutational profile with subsequent resections in chemo-refractory patients referred for ACT.

To accomplish this, we focused on patients with metastatic colorectal cancer enrolled in neoantigen-directed ACT protocols at our institution. In addition to sequencing freshly resected metastases and performing neoantigen screening, we sequenced matched primary FFPE samples to analyze neoantigen distribution between metastatic and primary tumors (figure 1A). As a secondary objective, we investigated whether mutations unique to primary tumors could be recognized by patients’ T cells, potentially broadening the neoantigen repertoire.

Figure 1

Mutation landscape in metastatic versus matched FFPE primary colorectal tumor samples. (A) Study design. (B) Number and proportion (inset) of shared versus unique non-synonymous mutations in metastatic (top) and matched FFPE primary tumor samples (bottom). Samples are ordered by the number of mutations detected in metastatic tumors. For patients who had >1 metastatic tumor sample analyzed, distinct mutations detected from all samples were counted. For those patients, an average number of mutations per metastatic tumor is indicated with a black line. FFPE primary tumors were analyzed as single regions for all patients. (C) Number of mutations in metastatic versus primary tumors. Each square represents one patient; lines in floating bars represent the cohort means. The analysis included all mutations, mutations affecting genes annotated in COSMIC-CGC database (“putative driver genes”), and mutations annotated in COSMIC-CGC as recurrent (“putative driver mutations”). (D) The proportion of all mutations shared between metastatic and primary tumor samples, as well as those unique to each tumor. (E) Comparing the proportion of shared mutations detected in metastatic versus primary tumors. Each square represents one patient; lines in floating bars indicate group means. Three classes of mutations correspond to those in (C). For (C) and (E), comparisons were performed using a paired Student’s t-test. Panel (A) was designed using www.BioRender.com. COSMIC-CGC, Catalog of Somatic Mutations in Cancer-Cancer Gene Census; FFPE, formalin-fixed paraffin-embedded.

MethodsPatients

The patients were evaluated at the Surgery Branch of the National Cancer Institute (NCI) for the management of relapsed or refractory metastatic colorectal cancer, under screening and cell harvest protocols approved by the NCI Institutional Review Board (NCT00001823 and NCT00068003, respectively). They had provided informed consent for enrollment in these studies. Eligible patients were subsequently enrolled in and treated under one of the cell therapy protocols (NCT01174121, NCT03745326 or NCT03190941).

Sample procurement and preparation

Metastatic tumor samples were obtained through surgical resection, and matched peripheral blood mononuclear cells (PBMCs) were collected via leukapheresis. Genomic DNA and RNA were extracted from both using the AllPrep DNA/RNA Miniprep Extraction Kit (QIAGEN, Germantown, Maryland, USA).

Primary tumor FFPE samples were procured from referring institutions following a review of pathology reports; specimens from single regions containing viable tumor were selected. These were received either as slides (10–20 microns thick) or as blocks, later sectioned onto slides. Every sixth slide was stained with H&E for review by an in-house pathologist. Areas containing tumor cells were marked and dissected using a scalpel; their DNA was extracted using the Covaris truXTRACT Kit and Covaris ME220 Focused Ultrasonicator (Covaris, Woburn, Massachusetts, USA).

Mutation calling

Tumor-specific single nucleotide variants (SNVs) and short insertions and deletions (indels) were identified from tumor samples using whole exome sequencing (WES) as previously described.7 Briefly, WES libraries were prepared from ~200 ng genomic DNA using SureSelectXT HS Target Enrichment System coupled with Human All Exon V6-V7 target bait (Agilent Technologies, Santa Clara, California, USA). Libraries were sequenced on a NextSeq 550 desktop sequencer using High Output Flow Cell V.2.5 (300 cycles) per manufacturer’s instructions (Illumina, San Diego, California, USA). Output from the sequencer was de-multiplexed and converted to FASTQ format using bcl2fastq program (Illumina). The reads were first trimmed using Trimmomatic19 and then aligned to human genome build 19 using NovoAlign MPI (http://www.novocraft.com/). Duplicates were marked using Picard’s MarkDuplicates tool; in/del realignment and base recalibration were carried out according to the GATK best practices workflow (https://www.broadinstitute.org/gatk/). After the data cleanup, pileup files were created using samtools mpileup (http://samtools.sourceforge.net). SNVs were called using Varscan2,20 Strelka,21 Sniper,22 and MuTect.23 The following criteria were applied: (1) ≥10 tumor and normal read counts (2) ≥7% variant allele frequency (3) ≥4 tumor variant reads, and (4) detected by ≥2 callers. Indels were called using VarScan 2 and Strelka, according to criteria (1–3). Finally, all variants were annotated using ANNOVAR (http://annovar.openbioinformatics.org). Mutations with ≥3 occurrences listed in the Catalog of Somatic Mutations in Cancer (COSMIC; https://cancer.sanger.ac.uk/cosmic) were labeled as recurrent.

Mutation clonality analysis

To determine the clonality of each mutation (ie, cancer cell fraction, CCF), local copy number and tumor purity were derived from WES using Sequenza,24 with matched normal samples as references and hg19 as coordinates. Then, mutation copy numbers were calculated by integrating local copy number, tumor purity and variant allele frequency (VAF).25 Only mutations with tumor coverage of ≥10× were used for this analysis. Finally, mutation copy number, tumor purity and VAF were integrated using PyClone.26 Mutations with read depth >4 and VAF>3% were clustered using PyClone V.1.3.0 Dirichlet process clustering, grouping clonal and subclonal mutations based on their CCF estimates. Mutations were deemed clonal if their 95% CCF CI overlapped with 1. Otherwise, they were considered subclonal. For this, PyClone was run with 50,000 iterations and a burn-in of 1,000.

Gene expression analysis

Raw sequencing data quality was assessed using the FastQC tool.27 Reads exhibiting poor quality (eg, short length, excessive “N” content) or low-quality 5’ and 3’ ends were filtered or trimmed, respectively, using FastP.28 Filtered paired-end FASTQ files were then analyzed using Kallisto29 to estimate transcript abundance. The latter was then converted to gene abundance using the “tximeta” R library.30

Screening libraries

Tumor-specific mutation sequences were used to generate tandem minigene (TMG) and peptide pool (PP) screening libraries. TMGs were constructed as described previously.3 Briefly, minigene sequences were generated first, encoding either non-synonymous point mutations (with the mutated amino acid flanked by 12 wild-type (WT) amino acids) or frameshift indels (with 12 WT amino acids preceding the new reading frame, and terminating at the next stop codon). Next, up to 24 minigenes were linked together to generate TMGs, which were codon-optimized, synthesized and ligated into a pcRNA2SL vector. Linearized vector DNA was in vitro transcribed into RNA using mMessage mMachine T7 Kit (Thermo Fisher Scientific Baltics UAB, Vilnius, Lithuania). RNA was purified using RNeasy Mini Kit (Qiagen, Germantown, Maryland, USA), quantified by spectrophotometry, and stored at −80°C until use.

PPs were composed of up to 20 crude-grade 25-mer peptides, corresponding to the minigenes (except for frameshift indels, which were represented by overlapping peptides). High-performance liquid chromatography (HPLC)-purified peptides were used in validation experiments. Peptides were synthesized in-house or obtained from GenScript (Piscataway, New Jersey, USA) as lyophilized powder, resuspended in dimethylsulfoxide and stored at –20°C until use.

Antigen presenting cells

Autologous B cells and monocyte-derived, immature dendritic cells (DCs) were used as antigen presenting cells (APCs). B cells were generated by purifying CD19+ PBMCs using magnetic microbeads (BD Biosciences, San Jose, California, USA). Purified CD19+ cells were cultured with irradiated NIH 3T3 CD40L cells in B-cell medium (Iscove's Modified Dulbecco's Medium [IMDM] supplemented with 2 mM L-glutamine, 25 mM 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid [HEPES], and Antibiotic-Antimycotic (Life Technologies, Carlsbad, California, USA), 10% human AB serum, and 200 U/mL interleukin (IL)-4 (PeproTech, East Windsor, New Jersey, USA)). Fresh medium was added on days 3–4; B cells were used fresh or cryopreserved on days 5–7.

DCs were generated using the plastic adherence method. First, cryopreserved PBMCs were thawed, washed, set to 5–10×106 cells/mL in AIM-V medium (Life Technologies), and incubated at 37°C in tissue culture flasks at approximately 106 cells/cm2. After 90–120 min, non-adherent cells were removed, and adherent cells were incubated with DC media (Roswell Park Memorial Institute [RPMI] medium with 2 mM L-glutamine, 25 mM HEPES, Antibiotic-Antimycotic (Life Technologies), 5% human AB serum, and 800 IU/mL granulocyte-macrophage colony-stimulating factor (GM-CSF) and 200 U/mL IL-4 (PeproTech)). DCs were detached from the flasks using a plastic scraper on days 4–6 and used immediately or cryopreserved until further use.

Tumor-infiltrating lymphocytes

TIL were generated from surgically resected tumors as described previously.31 Briefly, tumors were dissected free of hemorrhagic and necrotic areas and cut into 12–24 fragments measuring 1–2 mm in each dimension. These were plated in 24-well plates and cultured in 2 mL of complete medium (CM) with 6,000 IU/mL of IL-2 (Prometheus, San Diego, California, USA). CM consisted of RPMI medium, 2 mM L-glutamine, 25 mM HEPES, 10 µg/mL gentamicin (all from Life Technologies, Carlsbad, California, USA) and 10% human AB serum. Initial wells were processed as individual cultures, split in 1:2 fashion when fully confluent and, after approximately 6–8 weeks, cryopreserved until further use.

Assessment of neoantigen recognition by T cells

Putative neoantigen recognition was evaluated by co-culturing T cells (TIL, memory T cells from peripheral blood (PBL), or TCR-transduced T cells) with APCs overnight. Prior to co-culturing, APCs were either pulsed with peptides or transfected with TMG RNA. Peptide pulsing was performed with individual peptides (final concentration ~10 µM per peptide, unless indicated otherwise) or PPs (final concentration ~1 µM per peptide) at 37°C for 16–18 hours, followed by two washes.

For TMG electroporations, APCs were resuspended in Opti-MEM (Life Technologies) at 1×107 cells/mL and mixed with the TMG RNA in electroporation cuvettes (with either 2 or 4 µg of RNA added to the 50 or 100 µL of the cell suspension, respectively). Electroporations were performed using the BTX-830 electroporator (BTX Harvard Apparatus, Holliston, Massachusetts, USA) (150 V × 10 ms × 1 pulse for DCs or 150 V × 20 ms × 1 pulse for B cells). Electroporated cells were incubated overnight in an antibiotic-free medium and were washed once before the co-culture.

If previously cryopreserved, T cells were first rested for 24–48 hours in CM with IL-2 (6,000 IU/mL for TIL and 1,200 IU/mL for other T cells). On the day of the co-culture, they were washed twice to remove excess IL-2. For library screening, 0.2×105 of T cells were plated with 1×105 of peptide-pulsed or electroporated APCs into the MultiScreen-IP filter 96-well plates (MilliporeSigma, Burlington, Massachusetts, USA), which were precoated with interferon (IFN)-γ capture antibody (clone 1-D1K; Mabtech, Cincinnati, Ohio, USA). For other experiments, regular tissue culture U-bottom 96-well plates were used. Co-cultures were incubated overnight at 37°C and 5% CO2. Cell Stimulation Cocktail (phorbol 12-myristate 13-acetate (PMA) and ionomycin) (Affymetrix, San Diego, California, USA) was used as a positive control in 1:1,000 v/v ratio. The following day, T-cell activation was assessed by measuring IFN-γ production or upregulation of activation markers 4-1BB and OX40 on T-cell surface using flow cytometry, as described below.

In vitro sensitization of circulating memory T cells

Neoantigen recognition by PBL-derived memory T cells was assessed using a modified in vitro sensitization approach.13 T cells were first isolated from PBMCs using T Cell Enrichment Kit (BD Biosciences, San Jose, California, USA), followed by CD45RA-positive cell depletion using magnetic beads (also from BD Biosciences), yielding purified memory T cells. Next, 1–2×106 memory T cells were co-cultured (ie, stimulated) at a 4:1 ratio with DCs electroporated with TMGs or pulsed with PPs.

Co-cultures were initiated in 48-well plates, using media with IL-21 (30 ng/mL). On days 3, 6 and 9, cells were serially expanded into larger wells based on their growth, with fresh media containing IL-21 and IL-2 added to obtain final cytokine concentrations of 30 ng/mL and 300 IU/mL, respectively.

On day 12, expanded T cells were co-cultured with DCs or B cells loaded with the TMGs or PPs used for initial stimulation. The next day, T cells expressing the highest 4-1BB and OX-40 levels were sorted (see below for antibody selection) and subjected to rapid expansion. This was performed by mixing sorted T cells with >100-fold-excess of irradiated (6,000 rad) allogeneic PBMCs in 50/50 media (a 1:1 mix of AIM-V (Life Technologies) with 5% human AB serum and the CM) containing 30 ng/mL anti-CD3 antibody (OKT3; eBioscience, San Diego, California, USA) and 3,000 IU/mL IL-2. Cells were expanded for 10–14 days and then tested for the recognition of relevant TMGs or PPs. Concurrently, IVS with CEFX (JPT Peptide Technologies, Berlin, Germany), a viral peptide mix, was performed as a control.

Assessment of IFN-γ production

IFN-γ production from T cells was measured using enzyme-linked immunosorbent spot assay (ELISpot) or ELISA/ELISA-like method. For ELISpot, cell suspensions were harvested from the ELISpot plates and set aside for flow cytometry (see below). The plates were then processed as detailed previously,32 using biotinylated anti-human IFN-γ detection antibody (clone 7-B6-1) and streptavidin-ALP (both from Mabtech, Cincinnati, Ohio, USA). They were developed using the KPL BCIP/NBT substrate solution (SeraCare Life Sciences, Milford, Massachusetts, USA), and were scanned and counted using an ImmunoSpot plate reader (Cellular Technologies, Shaker Heights, Ohio, USA).

For the IFN-γ ELISA, co-culture plates were first spun at 300×g for 2–3 min at room temperature, followed by assaying the supernatants using an ELISA kit (Thermo Fisher Scientific, Waltham, Massachusetts). Cell pellets were resuspended in phosphate-buffered saline (PBS) with 0.5% fetal bovine serum (FBS) for flow cytometric analysis. ELISA plates were read on a SpectraMax 190 microplate spectrophotometer (Molecular Devices, Sunnyvale, California, USA) and analyzed using SoftMax Pro software (Molecular Devices). Alternatively, IFN-γ concentrations in the supernatants were measured with an electrochemiluminescence-based assay using the IFN-γ U-PLEX Kit (Meso Scale Diagnostics, Rockville, Maryland, USA). The plates were read on a MESO Sector S600 reader (Meso Scale Diagnostics) and analyzed using the accompanying software.

Flow cytometry

For all experiments, cells were stained with antibodies diluted in PBS/0.5% FBS in 1:50 V/V ratio at 4°C for 30 min. The antibodies were obtained from BD Biosciences (San Jose, California, USA): CD3 (clone SK7), CD8 (SK1), CD4 (SK3), CD62L (DREG-56), CD45RO (UCHL1), CD134 (OX40, clone ACT35), and CD137 (4-1BB, clone 4B4-1). Flow cytometry was performed on the FACSCanto I or II cell analyzer (BD Biosciences); sorting was performed on the SH800 sorter (Sony Biotechnology, San Jose, California, USA). Data was analyzed using FlowJo (TreeStar, Ashland, Oregon, USA).

Identification, synthesis and transduction of TCRs

To identify reactive TCRs from circulating memory T cells, single-cell TCR sequencing was performed using the Takara SMARTer Human scTCR a/b Profiling Kit—96 (Takara Bio USA, San Jose, California, USA), following the manufacturer’s instructions. Single cells upregulating 4-1BB following relevant peptide exposure were first sorted into 96-well plates, followed by complementary DNA (cDNA) synthesis and amplification using SMART technology with cellular barcoding. Subsequently, TCRA and TCRB transcript-encoding cDNA underwent further amplification and sequencing on an MiSeq instrument, employing paired-end 2×300 bp reads with the MiSeq Reagent Kit V.3 (600 cycle) (Illumina, San Diego, California, USA). Read extraction and clonality counts were determined using the MiXCR software package (V.2.1.12) (MiLaboratory, Russia). Identification of reactive TCRs from TIL was done as previously described.4

For TCR synthesis, TCR α and β constant regions were replaced with modified mouse constant regions to enhance TCR pairing and surface expression.33–35 TCRα and β chains were linked with a furin SGSG P2A linker,36 and then synthesized and cloned into an MSGV1 retroviral vector (GenScript, Piscataway, New Jersey, USA).

To transduce the TCRs into donor T cells, 293GP cells were first plated (1×106/well) in poly-D-lysine-coated 6-well plates (Corning, Tewksbury, Massachusetts, USA). On day 2, cells were transfected with 2 µg pMSGV1-TCR and 1 µg pRD114 (VSV-G) using Lipofectamine 2000 (Life Technologies). Concurrently, ~1–3×108 donor PBMCs were stimulated with 50 ng/mL anti-CD3 (eBioscience) and 1200 IU/mL of IL-2. On day 3, retrovirus-containing supernatants from 293 GP cells were transferred onto non-tissue culture 6-well plates that had been coated with RetroNectin (Takara Bio USA, Mountain View, California, USA). The plates were spun at 2000×g for 2 hours at 32°C and then seeded with stimulated PBMCs (2×106 cells/well) by centrifugation at 1000×g for 10 min at room temperature. Transduced cells were then incubated at 37°C in 50/50 medium with 1,200 IU/mL of IL-2 for an additional 5 days. Transduction efficiency was assessed by flow cytometry, using an anti-mouse TCRβ antibody (clone H57-597; Invitrogen, Waltham, Massachusetts, USA).

Statistical analyses

Statistical analyses were performed on GraphPad Prism V.10.2 software (GraphPad Software, La Jolla, California, USA). When applicable, data were expressed as mean±SD.

ResultsPatients and tumor samples

Archived primary tumor FFPE samples were sequenced for 22 patients whose metastatic tumors were resected at the NCI. Details on patient characteristics and NCI protocol procedures are provided in online supplemental table 1.

All patients presented with active metastatic disease, with a median of 2 affected organs. They had undergone 1–9 (median 4) different lines of systemic anticancer therapies, with four also receiving neoadjuvant therapy before primary tumor resection.

Between 1 and 4 (median 1.5) individual metastatic tumors were resected from each patient, primarily from the lung. These were fragmented to generate TIL cultures. For 7 patients, a single fragment underwent WES. For the remaining 15 patients, 2–8 fragments were sequenced, resulting in a total of 78 sequenced metastatic tumor samples. Of the 22 patients, 21 underwent TIL screening for cancer neoantigen recognition, as discussed below, with 12 receiving neoantigen-directed ACT.

Primary tumor FFPE samples were procured from referring institutions after metastatic tumor resection and a successful (positive) TIL screen for 10 patients (retrospective cohort). For the remaining 12 patients (prospective cohort), samples were collected and sequenced before TIL screening. The interval between primary and metastatic tumor resections ranged from 10.7 to 99.4 (median 36) months. Sample acquisition details are illustrated in online supplemental figure 1. Mutations identified in each patient’s metastatic and primary tumor samples are listed in online supplemental table 2.

FFPE primary tumors had fewer mutations than metastatic tumors

Across 22 patients, the number of non-synonymous somatic mutations, referred to here as tumor mutational burden (TMB), ranged from 61 to 391 (median 140.5) (figure 1B,C, and online supplemental table 3). The mutations in these samples, referred to as met-mutations, were predominantly SNVs, with fewer indels and stop-gain mutations. Their TMB did not correlate with clinicopathological characteristics such as resection site, number of prior therapies, or the interval between primary and metastatic tumor resection (not shown). However, there was a positive correlation between the TMB and the number of samples analyzed per patient (online supplemental figure 2A).

In individual metastatic samples, the TMB ranged from 61 to 198 (median 113) (online supplemental figure 2B and online supplemental table 4), with no correlation with the WES mean target coverage and only a weak correlation with estimated sample purity (online supplemental figure 2C). In 15 patients with >1 metastatic sample analyzed, 7.2–63.7% of mutations (median 33.6%) were unique to a single sample, indicating varying levels of intratumoral and intertumoral heterogeneity (online supplemental figure 2B,D, online supplemental table 4). The proportion of mutations unique to a single sample was lower for mutations affecting “putative driver genes”, that is, genes classified as tier 1 or tier 2 in the COSMIC Cancer Gene Census (CGC), which are annotated as proven or potentially cancer-relevant, respectively.37 This trend was especially pronounced for COSMIC-CGC gene mutations reported as recurrent in cancer, which are enriched for cancer driver mutations (referred to as “putative driver mutations”; online supplemental figure 2E).

Concurrent analysis of matched primary tumor FFPE samples revealed 17–115 non-synonymous somatic mutations (median 70) (figure 1B,C). TMB of these samples did not correlate with WES mean target coverage or estimated tumor purity, both of which were lower than in metastases (online supplemental figure 3A,B). There was no correlation between TMB and tumor site (left vs right colon) or differentiation status (not shown). However, significantly fewer mutations were found in the four patients who had received neoadjuvant therapies, despite similar tumor purity and target coverage (online supplemental figure 3C). While sequencing tumors with low tumor cell content, such as those arising after neoadjuvant therapy, may detect fewer mutations,18 a possible impact of tumor location could not be ruled out. Two of these four patients had rectal cancer (the only rectal cancers in the study), where fewer mutations are expected compared with right-sided but not left-sided colon cancers.38 However, the small number of cases limits conclusions about the impact of tumor location on TMB in our study.

Primary tumor TMB was significantly lower than that of metastatic tumors, both when considering all patient-specific mutations (figure 1B,C) and those detected in individual metastatic samples (online supplemental figure 4A). TMB in the two sample types did not correlate (online supplemental figure 4B). The discrepancy in TMB persisted, although to a lesser extent, when considering only mutations affecting putative driver genes, but disappeared when focusing solely on putative driver mutations (figure 1C). This is consistent with the previous finding that metastatic tumors exhibit limited driver mutation heterogeneity compared with matched primary tumors.39

Most mutations in FFPE primary tumors were shared with metastatic tumors

Out of all 3570 met-mutations from 22 patients, 34.3% were shared with primary tumors, while 65.7% were unique to metastatic tumors (figure 1D). The proportion of shared mutations increased for mutations in putative driver genes (48% of 273 mutations), and further increased for putative driver mutations (78.8% of 66 mutations) (online supplemental figure 4C).

For 78 individual metastatic samples, the proportion of mutations shared with primary tumors ranged from 2.8% to 89% (median 54%, online supplemental figure 4D). Their TMB was negatively correlated with the percentage of mutations shared with the corresponding primary tumors (online supplemental figure 4E), consistent with previous data.18 When >1 patient sample was analyzed, 0.3% of 1,039 mutations detected in only one sample were shared with primary tumors, while 49.8% of 1,718 mutations detected in >1 sample were shared (online supplemental figure 4F).

In primary tumors, 80.7% of the mutations were shared with metastatic tumors (figure 1D and online supplemental figure 4C). Across the 22 patients, primary tumors had a significantly higher fraction of shared mutations, both when considering all mutations and those in putative driver genes (80.7% vs 34.3% and 82.4% vs 48%, respectively) (figure 1E). However, the fraction of shared putative driver mutations was more similar between the two (91.2% vs 78.8%), consistent with previous reports.39 40

A total of 38 validated neoantigens were identified in metastases from 18 patients

TIL from the metastatic tumors of 21/22 patients were screened for recognition of cognate met-mutations, following a previously published protocol.4 For example, from patient 4371, who presented to the NCI with metastatic colon cancer affecting the liver and lungs, three distinct metastases were surgically removed from the lungs. WES of their representative fragments revealed 145 met-mutations, with 90 shared among all three tumors (figure 2A). Sequencing of the primary tumor FFPE sample detected 80 mutations, 13 of which were exclusive to the primary tumor, while 67 were shared with all three metastatic tumors.

Figure 2

Example of screening TIL for recognition of cancer neoantigens (patient 4371). (A) The number and distribution of all tumor-specific mutations detected in the primary formalin-fixed, paraffin-embedded sample and three lung metastases resected at the National Cancer Institute. (B) TIL cultures from 24 tumor fragments (F1-24) were co-cultured with autologous DCs that were either pulsed with ten 25-mer peptide pools (PPs) or transfected with corresponding TMGs representing 144 mutations detected in metastatic tumors, as well as 13 mutations detected exclusively in primary tumors. Names of PPs and TMGs containing the latter are underlined. A TMG encoding point mutations from another patient (Irrel. TMG) and DMSO were used as negative controls for TMG and PP testing, respectively. Results of IFN-γ ELISpot are depicted. (C) F13 TIL were co-cultured with DCs that were pulsed with individual peptides contained in PP9; names of genes harboring mutations exclusive to the primary tumors are underlined. Numerical suffixes represent different transcripts of the same genes or, in the case of indels, different overlapping peptides. PMA/ionomycin (PMA) was used as a non-specific positive control. Results of IFN-γ ELISpot and flow cytometric analysis of 4-1BB expression on the cell surface are depicted together. (D) 4-1BB upregulation was analyzed using flow cytometry following an overnight co-culture of F13 TIL with autologous DCs that were pulsed either with PP9 (red dots) or DMSO (black dots). Data was gated on live CD8+ T cells. Cells in the gated area were subjected to single-cell sorting and TCR sequencing. (E) T cells from an unrelated healthy donor were transduced with the PP9-reactive TCR and co-cultured with the patient’s DCs cells that were pulsed with decreasing concentrations of BIRC6P1505T (black squares) and corresponding wildtype peptide (empty squares). The 9-mer BIRC6P1505T peptide was identified from the 25-mer sequence as the minimal epitope in a separate experiment (not shown). IFN-γ production was measured the following day using IFN-γ ELISA. DCs, dendritic cells; DMSO, dimethylsulfoxide; ELISpot, enzyme-linked immunosorbent spot assay; IFN, interferon; PMA, phorbol 12-myristate 13-acetate; TCR, T-cell receptor; TIL, tumor-infiltrating lymphocyte; TMG, tandem minigene.

Next, TIL from the three metastases were screened for recognition of peptides and TMGs representing 144/145 met-mutations. Several reactive TIL were detected; the strongest responses included fragment 11 (F11) TIL recognizing PP5 and TMG5, F13 recognizing PP9, and F24 recognizing PP8 (figure 2B).

Subsequently, these TIL were retested with individual peptides from the reactive PPs or TMGs. For example, F13 TIL were co-cultured with peptides from PP9, revealing CD8+ T cell-mediated recognition of a peptide representing the BIRC6P1505T mutation (figure 2C). Single-cell RNA sequencing of sorted CD8+ cells that upregulated 4-1BB, a T-cell activation marker, in response to PP9 (figure 2D) identified a single TCR. T cells from an unrelated healthy donor transduced with this TCR exhibited dose-dependent production of IFN-γ only in response to the mutant but not the WT peptide, confirming BIRC6P1505T as a cancer neoantigen (figure 2E).

Using this approach, a total of 28 neoantigens from 15 patients were identified and validated in WT-versus-mutant experiments. They constituted a small fraction (1.08%) of the 2597 screened met-mutations, consistent with our previous findings.4 Details on these neoantigens and the screening process are summarized in online supplemental table 5.

To enhance the study, 10 additional neoantigens identified using alternative methods were also incorporated (online supplemental table 5). These included one neoantigen (SUN1A177T) identified by screening circulating PD-1+ lymphocytes,41 five neoantigens detected by testing TCRs identified through single-cell gene expression profiling of tumor digests,42 and four KRAS neoepitopes (one KRASG12D, two KRASG12V, and one KRASG13D) identified through targeted screening against mutant KRAS gene products.43

In total, 38 neoantigens from 18 patients were available for further analysis (herein referred to as “met-neoantigens”). Among them, 12 were encoded in putative driver genes. Seven of these represented known recurrent mutations in cancer: KRASG12D detected in three patients, KRASG12V detected in two patients, and KRASG13D and TP53R175H detected in one patient each. The remaining 26 neoantigens were not documented in COSMIC and thus appear to be unique to the patients’ tumors from this study.

Analysis of primary tumors captured 25/38 met-neoantigens

Out of 38 met-neoantigens, 25 (65.8%) were detected in matched primary tumor samples (figure 3), with similar detection rates for MHC class I-restricted (18/26, 69%) and class II-restricted neoantigens (7/12, 58%). 13 of 26 (50%) neoantigens encoded by genes not annotated as putative driver genes were detected in primary tumors (figure 3). Eight were clonal, three subclonal and two (MAGED2K150E in patient 4405 and SMU1M105L in patient 4371) had read counts below the standard mutation calling threshold (see Methods). The 50% detection rate was largely consistent with that of all screened mutations in non-putative driver genes (39.4% of 2,398 mutations) (online supplemental figure 5A).

Figure 3

Distribution of neoantigens in metastatic tumors compared with matched formalin-fixed, paraffin-embedded primary tumor samples. Left panel illustrates detection and clonality status for each of the 38 neoantigens that were identified in metastatic tumors (“met-neoantigens”) from 18 patients in this study. Patients are ordered by the number of separate metastatic tumor samples sequenced. Up to three separate tumors were analyzed (labeled as M (metastasis) 1–3); some of them were further divided and analyzed as separate fragments (labeled as F (fragment) 1–4). Right bottom panel provides a breakdown of met-neoantigen detection in primary tumors, according to the affected gene (annotated vs not annotated in COSMIC-CGC), mutation recurrence status (unique vs reported as recurrent in COSMIC-CGC), and MHC restriction. COSMIC-CGC, Catalog of Somatic Mutations in Cancer-Cancer Gene Census; MHC, major histocompatibility complex.

In contrast, all 12 neoantigens encoded by putative driver genes were detected in primary tumors, including five unique neoantigens (three detected as clonal and two as subclonal) and seven neoantigens encoding recurrent mutations, all detected as clonal (figure 3). Interestingly, while 100% of the neoantigens encoded by putative driver genes were detected in primary tumors, only 57.3% of the 199 screened met-mutations in these genes were detected (online supplemental figure 5A). This enrichment may be attributed in part to the over-representation of KRAS neoantigens identified through KRAS-specific screening and should be interpreted cautiously due to the relatively small sample size.

In 15 patients with multiregion or multitumor sequencing, only 0.3% of 635 screened met-mutations detected solely in one metastatic sample were shared with primary tumors (online supplemental figure 5B). The only validated neoantigen among these 635 was FUT1343_344del, identified in the metastatic but not the primary tumor of patient 4393 (figure 3). Conversely, 55.2% of the 1,327 screened met-mutations detected in ≥2 metastatic samples were shared with the primary tumors. Of these, 22 were immunogenic, and 14 (63.6%) were also present in the primary tumors. These findings suggest that currently used neoantigen assays are more likely to identify mutations present in multitumor samples. This observation, which merits further investigation, may be attributed to the higher likelihood of TIL detecting mutations found in multiple tumor fragments, as mutations confined to a single fragment may only be recognized by TIL from that specific fragment.

These results suggest that using primary tumor FFPE samples for screening could detect most (66%) met-neoantigens. None of the met-neoantigens absent from primary tumors were encoded in putative driver genes. Their predicted clonality did not differ from those detected in the primary tumor, nor did the expression of the genes harboring them (online supplemental figure 6). Additionally, in patients with multiregion or multitumor sampling, 8/9 met-neoantigens absent from primary tumors were present in all metastatic tumor samples sequenced (online supplemental figure 6C).

In summary, relying exclusively on FFPE samples for screening may result in the failure to detect some neoantigens that are clonal, highly expressed, and shared across metastatic tumor samples. The clinical relevance of targeting these neoantigens in colorectal cancer remains unclear, however, due to insufficient number of clinical responses to neoantigen-directed ACT.

Mutations exclusive to primary tumors were not recognized by TIL but were recognized by circulating memory T cells

For 10 patients from the prospective cohort, peptides and minigenes encoding mutations unique to primary tumors were screened using autologous TIL, alongside the met-mutations (figure 1A). For six of these patients, memory (CD45RA-negative) T cells from blood were also tested for mutation recognition (figure 4A).