Histopathologic and Molecular Biomarkers of PD-1/PD-L1 Inhibitor Treatment Response among Patients with Microsatellite Instability–High Colon Cancer

1. Patients and study group

Patients diagnosed with dMMR/MSI-H colorectal adeno-carcinoma and treated with ICIs at Asan Medical Center (Seoul, Republic of Korea) between October 2015 and February 2020 were screened for analysis. The patients were either treated with pembrolizumab or were enrolled in one of the phase 2 investigator-sponsored clinical trials of other ICIs, including NCT03150706 (avelumab for previously treated dMMR/MSI-H or POLE-mutant colorectal cancer) and NCT-03435107 (durvalumab for previously treated dMMR/MSI-H or POLE-mutant colorectal cancer) [9,10]. Patients were treated with the conventional ICI doses: intravenous pembrolizumab 200 mg every 3 weeks, avelumab 10 mg/kg intravenously every 2 weeks, and durvalumab 1,500 mg (20 mg/kg for patients with body weight ≤ 30 kg) intravenously every 4 weeks as described in each of the study protocols.

Among the screened patients, those who had finished the study treatment or had undergone disease evaluation once or more at the time of data collection (March 2020) were included in the study. The patients were then divided into two groups according to ICI treatment response (responder group vs. non-responder group). The responders were patients who received ICI treatment for > 4 months without progression after two consecutive disease evaluations within 8 to 9 weeks. The criteria were based on the results from a previous trial of pembrolizumab for pre-treated metastatic CRC patients. From the KEYNOTE-164 study, the median progression-free survival (PFS) was 4.1 months for dMMR/MSI-H metastatic CRC patients who received pembrolizumab after one or more prior treatments [11]. Patients who withdrew consent due to symptomatic deterioration before disease evaluation were included in the non-responder group.

2. Ascertainment of MSI status

Each patient was required to undergo a test for dMMR or MSI-H, or a next-generation sequencing (NGS) test for POLE mutation, which was also one of the eligibility criteria of the two trials. The tests for dMMR and MSI-H included MMR protein immunohistochemistry (IHC), polymerase chain reaction (PCR) fragment assay, and targeted NGS in which the MSI-H status was determined by a tumor mutational burden (TMB) ≥ 40 and an I-index (insertion/deletion mutation to whole mutation percentage) ≥ 9%, as previously described [12].

In current practice, the test methods for MMR and MSI-H are not standardized across patients, and previous studies have reported a considerable degree of discrepancy between MMR and MSI-H test results [9,13]. Therefore, to assure the MSI status of patients before biomarker analysis, we reviewed the results of IHC, PCR fragment assay, and targeted NGS. In cases of inconsistency among the results of IHC, PCR, and NGS, the MSI status was determined according to the results of a thorough review. NGS results were prioritized for determining the MSI status because our NGS testing based on TMB performs well even at low tumor cellularity (10% or more) compared with PCR testing, which requires a tumor cellularity > 20% to yield reliable results. IHC results are often affected by tissue quality, and misinterpretations of IHC results are known to be the most common cause of discrepancies between IHC and molecular testing [14]. Therefore, we prioritized NGS results over PCR and IHC results. For those without NGS results, a pathologist (J.K.) determined MSI status by reviewing the IHC and/or PCR results. Analyses of histopathologic characteristics and immune gene expression profiles were performed for each patient with a verified MSI-H status.

3. Clinical and histopathologic variables

Baseline clinical characteristics, including initial stage, previous treatments, mutational status of KRAS, NRAS, and BRAF, follow-up duration, and survival status, were obtained from the clinical trial database. Histopathologic features were evaluated, including histologic cancer subtypes, neutrophil infiltration grades, lymphocyte infiltration grades, tumor borders, Crohn-like lymphoid aggregate status, and lymphovascular invasion. Additional PD-L1 22C3 IHC (DAKO/Agilent, Santa Clara, CA) analyses were performed for cases with sufficient archival tissue for staining. PD-L1 results were interpreted by a pathologist (J.K.) by combined proportion score (CPS), defined as the ratio of all PD-L1–positive cells to viable tumor cells [15]. PD-L1 results were considered as positive if the CPS was ≥ 1.

4. IHC for PRAME

Differential expression of PRAME according to treatment response was examined using tumor tissues obtained during surgery for routine diagnostic pathologic examinations were analyzed with IHC for PRAME using anti-PRAME antibody (1:1,000, rabbit monoclonal, clone EPR20330, catalog No. ab219650, Abcam, Cambridge, UK). Briefly, 4-μm-thick sections of formalin-fixed, paraffin-embedded (FFPE) tissues were obtained with a microtome, transferred onto silanized charged slides, dried for 10 minutes at room temperature, and incubated at 65°C for 20 minutes. The tissue sections were processed by heat-induced epitope retrieval method using Cell Conditioning 1 buffer for 64 minutes and incubated for 32 minutes with the anti-PRAME antibody in a BenchMark XT automatic immunostaining device (Ventana Medical Systems, Tucson, AZ) according to the manufacturer’s instructions. Antigen-antibody reactions were visualized using the ultraView Universal Alkaline Phosphatase Red Detection Kit (Ventana Medical Systems). Counterstaining was performed using Ventana Hematoxylin II for 12 minutes and Ventana Bluing Reagent for 4 minutes. Finally, all slides were removed from the stainer, dehydrated, and coverslipped for microscopic examination. Slides in which > 1% of cancer cells were immunostained for PRAME were considered as positive for PRAME.

5. Differential gene expression and pathway analyses

Total RNA was extracted from the FFPE tissues of each patient. Quality control (QC) of each sample was performed using a Denovix DS 11 AATI Fragment Analyzer (Wilmington, DE) to evaluate the quantity and condition of the isolated RNA before analysis. Total RNA of approximately 100 μg was used for gene expression analysis, and the input amount of total RNA was increased for samples with excess RNA strand fragmentation. Immune-related gene expression profiling was performed using the NanoString nCounter platform (NanoString Technologies, Seattle, WA) with a PanCancer Immune Profiling Panel composed of 730 immune-related genes and 40 internal reference genes. The prepared RNA was thawed just before analysis and mixed with the reporter code set and probe set in a hybridization buffer. The hybridization process was performed at 65°C for 16 to 24 hours and then moved to a NanoString nCounter preparation station for cleansing of inadequately hybridized probes, and the properly hybridized transcript-probe complexes were immobilized on the cartridge. Finally, the fixed samples on the cartridge were scanned and read by the NanoString nCounter Digital Analyzer (NCT-DIGT-120) and recorded as reporter code count files, which were analyzed in nSolver software (NanoString Technologies) for the QC process, including image QC, binding density QC, as well as positive and negative control QC. The expression levels of each gene in the samples with adequate QC data were normalized in the nSolver software using a positive control and housekeeping genes. The immune cell type was annotated based on the annotation file provided by NanoString Technologies for the nCounter PanCancer Immune Profiling Panel.

From the normalized gene expression levels from the NanoString nCounter assay using the PanCancer Immune Profiling Panel, differential gene expression analyses were performed in responders and non-responders by comparing the normalized expression levels of each gene by Wilcoxon rank-sum test. The nominal p-values were initially adjusted according to the false discovery rate (FDR); however, all genes had an FDR of > 0.05 due to the small sample size. Therefore, we considered using fold change (FC) and the genes with nominal p-values < 0.05 and log₂-transformed fold change (log₂FC) > 0.5 or < −0.5 were considered as candidate genes. To identify the functional ontology of the candidate genes, we performed unsupervised hierarchical clustering and gene set enrichment analysis.

6. Predictive modeling using machine learning and internal validation

Random forest (RF), a machine learning classification modeling approach, was utilized using the Python package sklearn v0.24.1 to generate a predictive model for classifying patients into PD-1/PD-L1 blockade response groups based on the genes with significant differential expression. Validation of the predictive model was performed according to the following steps: Step 1: For the i^th sample (i = 1, …, n), divide the i^th sample from whole data as the training set and the remaining (n-1) patients as the validation set; Step 2: Apply classification models to the training set to fit a prediction model; Step 3: Apply the fitted prediction model to the validation set and calculate the predicted probabilities; Step 4: Repeat steps 1–3 for all n samples; Step 5: After completing the cross-validation, combine the predicted probability values of all samples calculated using the leave-one-out cross-validation (LOOCV) method. The overall accuracy was evaluated, and a single receiver operating characteristic curve was drawn, and the area under the curve (AUC) value was calculated.

7. Statistical analysis

For descriptive analysis of categorical variables, the chi-squared test or Fisher exact test was performed in R ver. 4.0.3 (R Foundation for Statistical Computing, Vienna, Austria), as appropriate. The Wilcoxon rank-sum test was used to evaluate the significane of differences in continuous variables between groups. Survival was estimated with the Kaplan-Meier method and was compared by log-rank tests using GraphPad Prism 5.0 (GraphPad Software, San Diego, CA) or R ver. 4.0.3. PFS was defined from the initiation of the study treatment until objective disease progression according to Response Evaluation Criteria in Solid Tumor v1.1 or death due to any cause, whichever came first. Multivariate logistic regression analysis was performed in the response group using the logistf v1.24 package in R. Multivariate Cox regression analysis was performed for PFS using the survival v.3.2–7 package in R. Two-sided p-values < 0.05 were considered statistically significant.

1. Patient screening and study design

A total of 50 patients who were enrolled and treated with PD-1/PD-L1 inhibitors between October 20, 2015, and February 27, 2020, at Asan Medical Center were screened. The median age was 59 years (range, 21 to 85 years), 40 patients (80%) were male, and all patients had Eastern Cooperative Oncology Group performance status of 0 or 1. Twenty-seven patients (54%) had initially metastatic disease at enrollment, while 23 patients (46%) had recurrence after surgical resection and adjuvant chemotherapy as needed. Nine patients (18%) received ICIs as the first-line regimen for palliative treatment, 18 patients (36%) as second-line therapy, and 23 patients (46%) as at least the third line. Among 33 patients with available tumor burden data, 17 (51.5%) had liver or lung metastases, and 26 patients (78.8%) had distant metastases elsewhere. Twenty-patients (40%) received pembrolizumab, while 13 (26%) and 17 patients (34%) received avelumab and durvalumab, respectively. The proportions of responders according to ICI were similar, although durvalumab was associated with the highest proportion of responders (9 patients, 52.9%) compared with pembrolizumab (8 patients, 40%) and avelumab (5 patients, 38.5%). At a median follow-up duration of 22.4 months, the median PFS was 3.7 months; 44% of patients (n=22) were categorized as responders (i.e., treated with ICIs for > 4 months without progression). The clinical characteristics of the 50 patients according to treatment response are summarized in S1 Table.

Among the 50 patients, 27 patients (15 responders vs. 12 non-responders), who were verified as having a dMMR/MSI-H tumor and adequate archival tissue, were included in this study (Fig. 1). At a median follow-up duration of 32.4 months, the median PFS of the 27 patients was 32.8 months, and the objective response rate was 44.4% (95% confidence interval [CI], 27.6 to 62.7). Immune-related gene expression analysis using NanoString was performed for 19 patients with dMMR/MSI-H (11 responders vs. 8 non-responders) after quality assurance of the tissue RNA.

Fig. 1

Case selection and study design according to the response to immune checkpoint inhibitor treatment among patients with deficient mismatch repair (dMMR)/microsatellite instability–high (MSI-H) metastatic colorectal cancers. CR, complete response; MSS, microsatellite stable; PD-1, programmed death-1; PD-L1, programmed death-ligand 1; PR, partial response; SD, stable disease.

2. Histopathologic determinants of PD-1/PD-L1 blockade response in dMMR/MSI-H CRCs

Histopathologic tumor features were compared among 27 patients with confirmed dMMR/MSI-H status according to treatment response (15 responders vs. 12 non-responders) (Table 1). The histologic CRC subtype distribution was significantly different between the two groups (p=0.003, Fisher exact test), with most of the patients with mucinous adenocarcinoma or signet-ring cell carcinoma (Fig. 2A) in the non-responder group. Compared with the non-responder group, patients in the responder group had abundant infiltration of immune cells, such as lymphocytes (p=0.001, Fisher exact test) and neutrophils (p=0.043, Fisher exact test) (Fig. 2A). We also found that the tumor border status was associated with treatment response, as an expansile tumor border (Fig. 2A) was significantly associated with treatment response (p=0.003, Fisher exact test); notably, none of the patients in the non-responder group had expansile borders. Interestingly, PD-L1 positivity and TMB were not significantly associated with the response to PD-1/PD-L1 blockade (Table 1). Also, there was no significant difference in the proportion of patients harboring RAS or RAF mutations between the two response groups (p=0.827, Fisher exact test). There were no significant differences in histopathologic features between dMMR/MSI-H tumors and microsatellite stable (MSS) tumors. TMB was significantly higher in dMMR/MSI-H tumors compared with MSS tumors, with median TMBs 104.7/Mb (range, 50.0/Mb to 176.0/Mb) and 12.5/Mb (range, 4.7/Mb to 17.2/Mb), respectively (p < 0.001, Wilcoxon rank-sum test) (S2 Table).

Table 1

Comparison of histopathologic characteristics among patients with dMMR/MSI-H according to treatment response

Fig. 2

Histopathologic features associated with ICI response among patients with dMMR/MSI-H CRC. (A) Representative histologic features of MSI-H CRCs among responders and non-responders. High degree of lymphocytic infiltration along the tumor border in the responder group (upper left, ×200). High degree of neutrophil infiltration grade along the tumor-stroma interface in the responder group (middle left, ×200). Expansile tumor border and surrounding of tumor cells by inflammatory cell infiltrates in the responder group (lower left, ×40). Infiltrative tumor border in the non-responder group (upper right, ×40). Mucinous adenocarcinoma with abundant extracellular mucin separating tumor cells from adjacent stroma in the non-responder group (middle right, ×100). Signet-ring cell carcinoma without signs of inflammatory cell infiltration along the tumor-stroma interface in the non-responder group (lower right, ×200). (B) Response status after ICI during follow-up and histopathologic features in patients with MSI-H CRCs. (C) Progression-free survival (log-rank test) of patients with MSI-H CRCs after ICI according to tumor histology, lymphocyte infiltration, and tumor border. (D) Multivariable logistic regression analysis of treatment response with histopathologic variables. CR, complete response; CRC, colorectal cancer; dMMR, deficient mismatch repair; ICI, immune checkpoint inhibitor; LA, lymphoid aggregate; MSI-H, microsatellite instability–high; ND, not determined; PD, progressive disease; PR, partial response; SD, stable disease.

In accordance with the initial responsiveness to immunotherapy, PFS was significantly associated with specific histopathologic variables (Fig. 2C, S3 Fig.), as mucinous adenocarcinoma and signet-ring cell carcinoma were associated with a significantly shorter PFS compared with conventional adenocarcinoma (p=0.004, log-rank test). Higher neutrophil infiltration grade (grade 2 to 3 vs. grade 0 to 1, p=0.016, log-rank test) and lymphocyte infiltration grade (grade 2 to 3 vs. grade 0 to 1, p < 0.001, log-rank test), presence of Crohn-like lymphoid aggregates (p=0.013, log-rank test), and expansile tumor border (p < 0.001, log-rank test) were associated with longer PFS. The presence of lymphovascular invasion was not associated with significant differences in PFS. In the multivariate logistic regression analysis, cancer histologic subtype (p=0.008) was independently associated with treatment response (Fig. 2D).

3. Differential expression of immune genes according to blockade responsiveness

Expression levels of the 730 immune genes in the PanCancer Immune Profiling Panel (NanoString Technologies) were compared according to treatment response. Immune cell type profiles were not significantly different between the two groups (S4 Fig.). At the individual gene level, 25 differentially expressed immune genes (p < 0.05, Wilcoxon rank-sum test and absolute log₂FC > 0.5) were identified (Fig. 3A, S5 Table). Using these genes, we performed pathway enrichment analyses to compare the activation of immune-related molecular pathways between the two groups. Responders had elevated activity in pathways, such as those yielding interferon-γ (IFN-γ) and regulation of immune effector process, whereas non-responders had elevated activity in pathways associated with phagocytosis, positive regulation of macrophage activation, and immunoglobulin/B-cell–mediated immune responses (Fig. 3B). When more stringent criteria were applied (p < 0.05, Wilcoxon rank-sum test and absolute log₂FC > 1.0), eight genes remained as differentially expressed (Fig. 3C), among which six genes (PRAME, CCL18, CXCL1, BST2, CXCL11, and CCL28) were specifically expressed in the responder group, and two genes (CD99 and ABCB1) were specifically expressed in the non-responder group (Fig. 3D).

Fig. 3

Differential expression analysis of immune genes between the two groups (11 responders vs. 8 non-responders). (A) Heatmap of differentially expressed immune genes (absolute log₂FC > 0.5 and p < 0.05 by Wilcoxon rank-sum test) between the two groups. (B) Enriched pathways in gene ontology enrichment analysis of responders and non-responders. (C) Volcano plots highlighting genes with significantly higher expression in the responder group and non-responder group (absolute log₂FC > 0.5 and p < 0.05 by Wilcoxon rank-sum test). (D) The comparative expression levels of genes with significantly higher expression in the responder group (BST2, CCL18, CCL28, CXCL1, CXCL11, and PRAME) and the non-responder group (ABCB1 and CD99). CRC, colorectal cancer; DEG, differentially expressed genes; MSI-H, microsatellite instability–high.

4. PRAME expression was associated with better response and prolonged survival

Of the six genes specifically expressed in the responder group, PRAME showed the highest FC (log₂FC=1.95). To examine the differential expression of PRAME, we performed IHC staining of PRAME (S6 Table). The IHC level of PRAME correlated well with the mRNA expression level from the NanoString panel, with a median normalized PRAME expression level of 76.6 (interquartile range, 66.8 to 146.4) among the PRAME-positive patients and 17.8 (interquartile range, 10.6 to 31.7) among the PRAME-negative patients (p=0.003, Wilcoxon rank-sum test) (Fig. 4A). All four PRAME-positive patients were in the responder group, and none were in the non-responder group (Fig. 4B and C). Moreover, except for one patient, all patients with a high mRNA expression of PRAME (i.e., higher than the median value) were in the responder group (Fig. 4C). High PRAME expression was associated with prolonged PFS compared with low PRAME expression (p=0.011, log-rank test) (Fig. 4D). Additionally, the prognostic significance of PRAME mRNA expression was independent of cancer histology (p=0.023, multivariate Cox regression) (Fig. 4E), thus showing that cancer histology was independently associated with ICI treatment response.

Fig. 4

Association between PRAME expression and response to ICIs. (A) Representative IHC results of PRAME with nuclear expression in CRC cells (×100) and significant correlation between protein expression and mRNA expression (p=0.0036, Wilcoxon rank-sum test). (B) Spider plot of the changes in the sum of target lesions from the baseline along ICI treatment with annotation of the PRAME IHC results. (C) Swimmer plot showing the clinical response and duration of ICI treatment with PRAME IHC results and PRAME mRNA expression levels. (D) Progression-free survival outcomes according to PRAME protein expression and PRAME mRNA expression among patients with MSI-H CRCs after ICI treatment (log-rank test). (E) Multivariable Cox regression analysis for progression-free survival. CI, confidence interval; CR, complete response; CRC, colorectal cancer; ICI, immune checkpoint inhibitor; IHC, immunohistochemistry; MSI-H, microsatellite instability–high; PD, progressive disease; PR, partial response; SD, stable disease.

5. Predictive modeling of PD-1/PD-L1 blockade response based on immune-related gene expression

The PD-1/PD-L1 inhibitor response prediction model using an RF algorithm was built based on the top eight immune-related genes with differential expression, and the prediction model was validated using the LOOCV method (Fig. 5A). In accordance with the aforementioned finding of the association between PRAME expression and treatment response, PRAME had a high rank in terms of feature importance among the eight genes (Fig. 5B). The accuracy of the response prediction model was 0.84 (95% CI, 0.60 to 0.96), with a cross-validated AUC of 0.93, sensitivity of 0.91, specificity of 0.75, a positive predictive value of 0.833, and a negative predictive value of 0.857 (Fig. 5C).

Fig. 5

Prediction modeling for treatment response to ICIs among patients with MSI-H CRCs. (A) Overview of the processes of prediction model building using RF based on immune-related gene expression and internal validation of the model using the LOOCV method. (B) Feature importance of input genes in modeling by RF. (C) Performance of the prediction model. AUC, area under the curve; CRC, colorectal cancer; DEG, differentially expressed genes; ICI, immune checkpoint inhibitor; LOOCV, leave-one-out cross-validation; MSI-H, microsatellite instability–high; NPV, negative predictive value; PPV, positive predictive value; RF, random forest; ROC, receiver operating characteristic.