Clonality and antigen-specific responses shape the prognostic effects of tumor-infiltrating T cells in ovarian cancer

CD8+ tumor-infiltrating lymphocytes (TILs) are not all specific for tumor antigens, but can include bystander TILs that are specific for cancer-irrelevant epitopes, and it is unknown whether the T-cell repertoire affects prognosis. To delineate the complexity of anti-tumor T-cell responses, we utilized a computational method for de novo assembly of sequences from CDR3 regions of 369 high-grade serous ovarian cancers from TCGA, and then applied deep TCR-sequencing for analyses of paired tumor and peripheral blood specimens from an independent cohort of 99 ovarian cancer patients. Strongly monoclonal T-cell repertoires were associated with favorable prognosis (PFS, HR = 0.65, 0.50–0.84, p = 0.003; OS, HR = 0.61, 0.44–0.83, p = 0.006) in TCGA cohort. In the validation cohort, we discovered that patients with low T-cell infiltration but low diversity or focused repertoires had clinical outcomes almost indistinguishable from highly-infiltrated tumors (median 21.0 months versus 15.9 months, log-rank p = 0.945). We also found that the degree of divergence of the peripheral repertoire from the TIL repertoire, and the presence of detectable spontaneous anti-tumor immune responses are important determinants of clinical outcome. We conclude that the prognostic significance of TILs in ovarian cancer is dictated by T-cell clonality, degree of overlap with peripheral repertoire, and the presence of detectable spontaneous anti-tumor immune response in the patients. These immunological phenotypes defined by the TCR repertoire may provide useful insights for identifying “TIL-low” ovarian cancer patients that may respond to immunotherapy.


Statistical methods
Examining the distribution of clone frequencies, we used Fisher's model for unseen species assuming an exponential model (shape = 1) to calculate the sequencing coverage; we estimated that 78% of tumor samples have at least 90% coverage and the median coverage is 83%. Low coverage was not associated with any of the repertoire statistics.
Supplementary Figure 1 and Supplementary Table 2 contain a spread of statistics meant to capture different features of the repertoire. We report on clonality as defined above as well as the TOPX statistics (the % of the repertoire captured by the most common X reads), the NX statistics (the number of clones required to capture X% of the repertoire) and the RX statistics (the % of the PBMC repertoire that maps to the corresponding clones identified by NX).

Patient characteristics and survival data
Study outcomes included overall survival (OS), the time until death; and progression-free survival (PFS), the time to recurrence or progression, both measured from the time of definitive surgery. Progression or recurrence was defined by time to radiologically confirmed relapse of disease. If no imaging data were available, progression was determined by increase in CA125 to twice the upper limit of normal or two times the nadir value if CA125 was never normal during primary treatment.

Sequencing variation and TCR repertoire statistics depend on degree of infiltration
Each patient's TIL TCR repertoire contained between 224 and 94,770 (1Q: 2,918; 3Q 15,190; median: 11,590) unique AA sequences (Supplementary Table 2). The number of reads in the tumor specimens (17,368,648) varied over three orders of magnitude and was associated with density of CD3 + T cell infiltration measured by IHC (Spearman's rho = 0.340, p = 0.001). Between 65.3% (1st quartile) and 84.0% (3rd quartile) of reads are captured by the most common 1000 clones implying that most samples have a large number of rare reads.
To leverage the deep TCR sequence and frequency information in our 99 samples, we first analyzed the variety of TCR sequences across all patients. We obtained 992,791 unique TIL TCR sequences in 289,963,375 normalized reads in combined 99 patients. In matched PBMC samples, there were 6,275,193 unique clones in 400,143,279 normalized reads. We found that, on average, two patients share 2.6% of PBMC clones (range: 0.5-4.7%) which is lower than previously reported results.
Noting that there were similar overall levels of reads, the overall TIL repertoire appeared more restricted than the PBMC repertoire. The most shared TIL clone was found in only 41 of the 99 patients while the most shared PBMC clones appeared in more than 90 of the patients. Again these clones did not comprise an unusually large fraction of the reads.
With respect to the usage of specific clones in the repertoire, the average rate of reads to their unique TCR clone may describe the evenness. For example, the mean ovarian cancer TIL repertoire had one unique clone for every 364 reads (median 1 per 199 reads). The range varied from 1 per 2714 reads (strongly restricted) to 1 per 25 reads, which was close to the noise threshold at 20 reads.

Individual TIL repertoires
We describe each TIL repertoire by plotting the relative frequency of each clone in order from most common to least using the log-rank to emphasize the most common clone and to de-emphasize the long tail of low abundance clones. All the plots in Supplementary Figure 2 have the same x-axis so the longer the right tail, the more unique clones in the repertoire. Additionally, we have split the cohort into roughly equal sized bins based on the TOP1 statistic. Therefore the more monoclonal cases and the more polyclonal cases are grouped together and the set of putatively oligoclonal cases are in between. Repertoire specific statistics are given for each patient.