Clinical responses to PD-1 inhibition and their molecular characterization in six patients with mismatch repair-deficient metastatic cancer of the digestive system

We report on our first six consecutive patients for whom treatment with checkpoint inhibitors was initiated between June 2016 and August 2017. Patients suffered from progressive MSI-H/dMMR metastatic cancer of the digestive system (four patients with colon carcinoma, one patient with duodenal carcinoma, one patient with cholangiocarcinoma). All patients had received at least one prior therapy and had evidence of progressive disease prior to checkpoint inhibition. All patients were naïve to anti-PD-1, anti-PD-L1 and anti-PD-L2 antibodies. Molecular analysis by targeted next generation sequencing (127 gene panel, 0.8 Mb) was performed post-hoc. Treatment-naïve tumor tissue of the primary tumor (patients P1 to P4 and P6) or a metastatic lesion (P5) was used for molecular testing. The study procedure was approved by the Medical Ethics Commission II of Heidelberg University (Medical Faculty Mannheim; 2020-807R) including a waiver for informed consent.

Patients received either pembrolizumab (2 mg/kg every 3 weeks, maximum dose 200 mg) or nivolumab (3 mg/kg every 2 weeks). Treatment was generally continued until unacceptable toxicity, or disease progression. Serum biomarkers (CEA, CA19-9, CA72-4) were measured on an individual basis but generally at baseline and if elevated at baseline further monitored along with radiographic assessments. Radiographic assessments were performed every two to four months depending on patient performance and disease dynamics.

Formalin-fixed paraffin-embedded (FFPE) tumor tissues were collected from the archives of the Institutes of Pathology in Mannheim, Bad Mergentheim and Speyer (all Germany). Histology was reviewed by two pathologists (DH, TG) and tumor areas containing at least 40% tumor cells were marked for molecular testing. Treatment-naïve tumor tissue (primary tumor for patients P1–P4 and P6, metastatic lesion for patient P5) was used for molecular testing.

DNA extraction of FFPE tumor and respective normal tissues was done as published previously (Hirsch et al. 2012). DNA concentration was measured by fluorometric quantitation (Qubit 3.0 Fluorometer, Life Technologies, Thermo Fisher Scientific, Carlsbad, CA, USA) using the Qubit dsDNA HS (High Sensitivity) Assay Kit (Life Technologies).

Mismatch repair/microsatellite status of tumors was determined by immunohistochemistry (IHC) and/or polymerase chain reaction (PCR) as described previously (Hirsch et al. 2018). Briefly, IHC was performed using the following primary antibodies: MLH1 (1:25; clone ES05, cat # M3640, Dako, Agilent Pathology Solutions, Agilent, Santa Clara, CA, USA), MSH2 (ready-to-use; clone FE11, cat # IR085, Dako), MSH6 (ready-to-use; clone EP49, cat # IR086, Dako), and PMS2 (1:50; clone EP51, cat # M3647, Dako). Detection was done using the EnVision Detection System, Peroxidase/DAB, Rabbit/Mouse (cat # K5007, Dako). IHC stainings were validated by internal and/or external positive controls as well as negative control specimens. IHC stainings were evaluated by two pathologists (DH, TG). Microsatellite PCR of tumor and corresponding normal DNA was done using a panel of five mononucleotide markers (BAT25, BAT26, NR-21, NR-24, and MONO-27; cf. MSI Analysis System, Promega), and a panel of two mononucleotide (BAT25 and BAT26) and three dinucleotide markers [D5S346, D2S123, and D17S250; so-called Bethesda panel; (Boland et al. 1998)]. Tumors were classified as MSI-H when two or more markers of either the Bethesda panel or the Promega panel showed an allelic size variation (i.e., a band shift compared with corresponding normal DNA).

Targeted next generation sequencing (NGS) was done using the xGen Pan-Cancer Panel (v1.5; Integrated DNA Technologies, Coralville, IA, USA), spanning 0.8 Mb of the human genome and targeting 127 significantly mutated genes implicated across 12 tumor tissues (Kandoth et al. 2013). We determined the sequences of the tumors and matched normal (non-tumor) DNA for our six patients (matched normal DNA was not available for P5). Library preparation on FFPE isolated DNA was done using the Nextera DNA Exome library preparation kit (Illumina, San Diego, CA, USA) with 50 ng according to the manufacturer’s protocol, except for 14 instead of 10 PCR cycles. The resulting paired-end libraries were pooled using 100 ng of each library. The library pool was used for targeted capture with the xGen Pan-Cancer panel according to the manufacturer’s protocol except for 12 instead of 10 PCR cycles. An aliquot of 1.4 pM was sequenced on a NextSeq500 system (Illumina) using the 150 cycle mid-output kit (2 × 76 bps). Alignment was done using BWA mem 0.7.12-r1039 (Li and Durbin 2009) to hg19. The mean read depth for the targeted regions (mean coverage) was 1061X. We minimized calling FFPE artifacts by applying a minimal variant allele frequency (VAF) threshold of 10% (Melendez et al. 2018). Details on the bioinformatic analysis and variant calling are provided in the Supplementary Information.

TMB was calculated as the number of somatic coding mutations per megabase (Mb), including non-synonymous (missense) mutations, synonymous mutations, nonsense (stop) mutations, and/or frameshift mutations present above 10% VAF after filtering. Non-coding alterations and mutations predicted to be germline were not counted. We compared our data to the TMB thresholds suggested by Fabrizio et al. (Fabrizio et al. 2018) (≥ 11.7 mutations per Mb, synonymous and non-synonymous mutations) and Schrock et al. (Schrock et al. 2019) (proposed cut-off 37.4 mutations per Mb with a range of 37–41 mutations per Mb, synonymous and non-synonymous mutations), which both are based on the F1CDx Foundation medicine assay (324 genes, 1.11 Mb) and a VAF of 5% or greater (Frampton et al. 2013). Furthermore, influence of keeping or removing COSMIC listed variants on TMB estimation was evaluated. To assess the impact of VAF on TMB, we tested different cut-offs of 5%, 7.5%, 10%, 15%, and 20%, respectively. Impact of germline filtering was compared between (1) only computational germline filtering, (2) filtering against the matched normal sample, (3) filtering against the panel of other, non-matched normal samples (n = 4), and (4) filtering against the matched normal sample and a local panel of normal samples (Kiel normal samples; n = 55). Normal samples used as reference for germline filtering were processed in the same way as the tumor samples.

To complement immunohistochemical and PCR-based MSI testing, we also assessed the microsatellite status from NGS data by applying MSIsensor on targeted NGS data from tumor and corresponding normal samples (Niu et al. 2014).

Response and progression were evaluated using Response Evaluation Criteria in Solid Tumors (RECIST) 1.1 criteria (Eisenhauer et al. 2009; Schwartz et al. 2016). Toxicities were graded based on the National Cancer Institute Common Terminology Criteria for Adverse Events CTCAE, version 5.0, published by the National Cancer Institute/National Institutes of Health in November 2017). Progression-free survival (PFS) and overall survival (OS) rates were calculated using the Kaplan–Meier method. PFS was defined as the time from the date of the initial dose of immune checkpoint inhibition to the date of disease progression or the date of death due to any cause, whichever occurred first. PFS was censored on the date of the last evaluable tumor assessment documenting absence of progressive disease for patients who were alive and progression-free. OS was defined as the time from the initial dose due to death of any cause. For patients who were still alive at the time of analysis, the OS time was censored on the last date the patients were known to be alive. Duration of response was the time of first response as defined by RECIST 1.1 to the time of disease progression and was censored at the last evaluable tumor assessment for patients who had not progressed. Analyses and graphs were generated with GraphPad Prism software (version 8.4.2; San Diego, CA, USA).

Our patient cohort consisted of six patients with mismatch-repair-deficient metastatic cancer of the digestive system (Table 1, Supplementary Table S1). Mismatch repair deficiency was assessed by polymerase chain reaction and/or immunohistochemistry. Cancers comprised three different entities: colorectal adenocarcinoma (4 patients), duodenal adenocarcinoma (1 patient) and cholangiocarcinoma (1 patient). All tumors were diagnosed between 2013 and 2016 in either locally advanced or metastatic stages (UICC stages IIIA-IV). Median age at diagnosis was 50 years (range 37–61). Four of the patients had a positive family history for gastrointestinal cancer and Lynch syndrome diagnosis was confirmed by genetic testing. Before immune checkpoint inhibitor therapy, all patients had received at least one prior line of chemotherapy and presented with metastatic progressive disease. PD-1 inhibition (nivolumab, pembrolizumab) for all patients was started between June 2016 and August 2017. Four patients were treated with pembrolizumab, while one patient received a combination of radiotherapy and nivolumab, and one patient nivolumab alone. Median time under PD-1 inhibition was 17 months (range 4–40). At the data cut-off, PD-1 inhibition is still ongoing in two patients, three patients had discontinued therapy due to excellent response with ongoing remission, and one patient had deceased from tumor progression. Median duration of response was 25 months (range  18–35 months). Patients were followed for a minimum of 25 months (median 27, maximum 40).

Responses to PD-1 inhibition were evaluated radiographically in all six patients based on RECIST v1.1 criteria (Fig. 1a, Supplementary Figure S1A). In addition, biomarker levels for CA19-9, CA72-4 and/or CEA were monitored over the course of treatment if they had been elevated at baseline (Fig. 1b, Supplementary Figures S1B-D). Objective radiographic responses (i.e., partial or complete response based on RECIST v1.1 criteria) were noted in all six patients, with two patients achieving a complete response after 2.4 (P3) and 18.5 months (P6), respectively. The time to first objective radiographic response ranged from 1.5 to 8.6 months (median 2.3 months). In three patients (P1, P2, P4) tumor marker level reduction preceded objective radiographic response. However, the progression of patient P5 at 31.8 months after treatment initiation was first indicated by a radiographic increase in tumor size.

Treatment was discontinued in three of six patients due to excellent response to anti-PD-1 therapy after 7.8 months (P2), 3.8 months (P3), and 4.9 months (P6), respectively, despite some residual disease by imaging in patients P2 and P6 at the time of treatment discontinuation (Fig. 1C and Fig. 2). As of the data cut-off, the time off therapy in these three patients was 31.1 (P2), 22.5 (P3) and 20.3 (P6) months, respectively. None of these three patients has shown evidence of cancer recurrence or progression since discontinuation of pembrolizumab or nivolumab so far. Instead, patient P6, who had partial response when taken off therapy, converted to complete response 13.6 months after treatment cessation.

Immunotherapy was continued in the other three patients whereby the dose for patient 1 was reduced to 4-weekly cycles after 13.3 months on treatment. Remarkably, the two disease progressions in our cohort (P1, P5) were observed in the patient group that continued with PD-1 inhibition. Patient 1 had the first evidence of progressive disease by imaging 27.1 months after starting PD-1 inhibition. Disease progressed rapidly and despite salvage therapy with FOLFOX, the patient passed away within one month after diagnosis of progressive disease. Patient 5 also had a progressing paraaortic lesion after 31.8 months of immunotherapy. The progressing metastatic lesion was excised by surgery and as of the data cut-off, which was 4.9 months after surgery, the patient is without tumor activity and PD-1 inhibition is continued. None of the patients had primary resistance to PD-1 inhibition, however, acquired resistance was noted in two patients, who developed progressive disease after an initial objective response to pembrolizumab. While one patient (P5) could be treated with local therapy (surgery) and survived and as of the data cut-off continues treatment with pembrolizumab, the other patient (P1) experienced rapid disease progression and died within one month after the first evidence of disease progression by imaging. Taken together, median progression-free survival (PFS) in our small cohort of six patients was 31.8 months (Supplementary Figure S2A); median overall survival (OS) has not yet been reached (median follow-up time of 29.9 months, range 25.2–40.2 months) (Supplementary Figure S2B). The estimates of PFS and OS at 1 and 2 years (12 and 24 months, respectively) were 100% (6 of 6 patients), respectively. The PFS and OS were not strikingly different in patients with colon cancer relative to those with the other GI cancer types. Neither PFS nor OS seemed to be influenced by tumors associated with Lynch syndrome. Adverse events during treatment were manageable and resembled those reported in other clinical studies using PD-1 inhibition (Le et al. 2015). No grade 3 or 4 events were noticed, in particular no immune-related adverse events leading to discontinuation of therapy were observed. Of note, patient P2 developed a reactive thymic hyperplasia under immunotherapy, which has still been present in the latest CT scan (Supplementary Figure S3). Detailed clinical history for each patient is provided in the supplement.

In an attempt to better understand the genetic basis of the good responses to PD-1 inhibition in our small series of six patients, we performed targeted sequencing with a 127 gene panel spanning 0.8 Mb, which is considered suitable for TMB assessment according to current recommendations (Buttner et al. 2019). In addition to non-synonymous (missense), nonsense (stop) and frameshift mutations, we included synonymous mutations into our TMB calculation, following the rationale of Chalmers et al. (2017). Chalmers et al. (2017) reasoned that synonymous mutations, though they are not likely to be involved in creating immunogenicity, are a signal of mutational processes that will also have resulted in non-synonymous mutations and neoantigens elsewhere in the genome. Immunogenicity of particular mutation types and which mutation types should be included into TMB calculation is still a matter of debate and no uniform method exists. When using gene panels biased toward genes with functional mutations in cancer for TMB calculation, exclusion of mutations listed as known somatic alterations in COSMIC (Bamford et al. 2004) has been suggested (Chalmers et al. 2017). Overall, our data shows that MSI-H/dMMR leads to an increased TMB in all our patients, though to a different extent. When including COSMIC listed mutations, mutation counts averaged to 170 mutations per Mb with a range of 42–397 (Fig. 3). When excluding COSMIC listed mutations, an average of 112 mutations was detected per Mb with a range of 20–208. Hence, exclusion of COSMIC listed variants decreased absolute TMB levels by 19–53%, respectively. However, relative TMB levels of samples to one another remained similar.

The impact of included mutation types and COSMIC variants is critical, in particular with respect to patients with borderline TMB like patient P6 in our cohort, who despite borderline TMB had a durable response and clinical benefit from PD-1 inhibition via nivolumab, which is still ongoing after 25.2 months. In general, TMB levels did not associate with complete or partial response in our small subset of six patients. Moreover, patient P6 achieved complete response after 18.5 months, despite having the lowest TMB in our cohort. The other complete responder, patient P3, had an intermediate TMB level in our cohort. Apart from particular mutations that are included or not, TMB is influenced by a plethora of other factors (Stenzinger et al. 2019, 2020) including but not limited to minimum VAF threshold and method of germline filtering. As expected, the higher the VAF threshold, the lower the mutation count (Supplementary Figure S4). For VAFs below 10%, the method of germline filtering had a considerable impact on mutation counts. If only in silico germline filtering was performed, mutation counts were remarkably higher at VAFs of 5% and 7.5% compared to filtering against a matched normal sample or a local panel of normal samples (Supplementary Figure S4). This implies that thresholds below 10% may not be advisable for samples with solely computational germline filtering due to technical sequencing artifacts and FFPE artifacts, in line with Melendez et al. (2018). As filtering against a local panel of normal samples generated using the same workflow could at least partially resolve this issue, this should be considered as a practical option for routine molecular diagnostics where additional sequencing of a matched normal sample is often not feasible. Of note, for patient 1, who showed rapid progression, no potential genomic target causing this could be identified by the NGS analysis of the primary tumor sample. Due to rapid progression, tumor material from the progressing lesions could not be obtained.

To evaluate whether MSI status can reliably be assessed by panel sequencing, we applied the MSIsensor software tool (Niu et al. 2014) to our dataset (Supplementary Figure S5). This revealed an MSIsensor score highly suggestive for MSI-H in all five patients, for which a matched normal sample was available, a prerequisite for this tool. From a technical point of view, NGS data nicely confirmed MSI based on the MSIsensor score, making it a potentially useful tool for future studies and clinical purposes.