Clinical Immunoprofiling Reveals that High Numbers of CD8⁺ and PD-1⁺ Cells Predict Superior Patient Survival Across Major Cancer Types Independent of Major Risk Factors

Clinical workflow and validation

ImmunoProfile was validated within a CLIA-certified molecular diagnostics laboratory (Figure 1A, Supplementary Figure 1, Supplementary Methods). FFPE tissue blocks from consented patients (DFCI 11-104; 17-000) were reviewed to select the “best block” for analysis. An H&E-stained slide was used to confirm tumor content. An additional slide was used for automated multiplex immunofluorescence (mIF) staining with a Leica BOND Rx and imaged on an Akoya PhenoImager at 200X magnification.^{30, 33, 34} A minimum of three and a maximum of six regions of interest (ROIs) were selected per whole slide image (WSI) with Inform software (version 2.4.8) according to pre-determined metrics to ensure the capture of immunological “hot spots” (if present). Technicians trained a random-forest-based machine-learning (ML) algorithm in Inform with representative cells and then used it to quantify defined cell types across ROIs. Custom tools (https://github.com/jason-weirather/pythologist) were used to calculate CD8⁺, PD-1⁺, CD8⁺PD-1^+, and FOXP3⁺ cell densities, PD-L1 tumor proportion scores (TPS), and PD-L1 combined proportion scores (CPS, Supplementary Methods).^{35, 36} An expert pathologist reviewed and signed out each case.

• Download figure
• Open in new tab

Figure 1.

A: Schematic outlining the ImmunoProfile workflow in the CLIA-certified laboratory. Consented patients and their diagnostic pathology material were first identified. Slides were cut to generate an H&E-stained slide, which a pathologist reviewed and confirmed to contain diagnostic material. An additional slide was stained on an autostainer and scanned, and a multiplex digital image of the stained tissue was generated. Technicians digitally annotated the images, including selecting regions of interest (ROI) and demarcating the tumor border under pathologists’ review and approval. This was followed by algorithmic training to identify positive and negative staining cells for each biomarker and automated quantitative scoring using image analysis software. A pathologist approved and signed out the image analysis results and the final scores. Results were reported to a clinically accessible database. Images and image analysis results were retained in a HIPAA-compliant data warehouse. Further details are provided in Supplementary Methods. B: The 2,023-patient pan-cancer ImmunoProfile cohort. The cases are separated into 14 major diagnostic categories and “other” rare cancer types. The number of cases within each major cancer type is further annotated by the tissue acquisition site (primary versus metastatic) and patients’ clinical stage. Further details are provided in Supplementary Materials. C: Violin plot showing the distributions of intratumoral immune cell density (log₂ of cells per mm²) for all cases, separated according to cancer type and organized from the highest to lowest average CD8⁺ immune cells. The blue lines indicate the immune cell densities at the 75^th and 25^th percentiles for the pan-cancer cohort. The black lines represent the mean value, the red boxes represent the 75^th and 25^th percentile thresholds, and the black outlines represent case numbers at the indicated density for each cancer type. Note that there are cases within the highest and lowest quartiles for each biomarker for each cancer type.

Clinical Annotation and Survival Analysis

Patient information was obtained from medical records and the DFCI OncDRS data warehouse with IRB approval.³⁷ Appropriate clinical experts also manually reviewed every case (Supplementary Table 1).

Statistical Methods and Prognostic Models

Comparisons between groups for continuous scaled data used Wilcoxon rank-sum or Kruskal-Wallis tests. Categorical comparisons used Fisher’s exact or chi-squared tests. Statistical significance was p ≤ 0.05. Overall survival (OS) was determined from the date of tissue acquisition and death from any cause. The follow-up of patients alive at data retrieval was censored at the date of last contact. We divided patients by their respective biomarker scores into tertiles when there were >100 cases and >50 deaths for a cancer type. We split patients by their respective biomarker scores at the median for smaller cohorts.

OS was estimated using the Kaplan-Meier method and compared using log-rank testing; 95% confidence intervals were calculated using log(-log) methodology. Median follow-up was based on Kaplan-Meier estimation with an inverted censor.

Associations between biomarker measurements and OS were assessed using Cox proportional hazards models adjusted for clinical risk factors selected before analyses, including cancer type (for the pan-cancer analysis) and the American Joint Committee on Cancer (AJCC) stage.³⁸, age, alcohol use, smoking, and sex. To avoid overfitting, the number of risk factors included in each model was adjusted according to cohort size and deaths before analysis. Multivariable models were fit for cancer types with > 30 deaths only. Immune cells with adjusted Wald p-values ≤ 0.1 from the single marker analyses were considered candidates in the multivariable setting (see Supplementary Methods).

ImmunoProfile reliably quantifies key immunological biomarkers in clinical samples

We modified a research protocol³⁵ to develop a new clinical workflow, ImmunoProfile, which combines automated mIF tissue staining, digital WSI capture, and ML-directed analysis to calculate CD8⁺, PD-1⁺, CD8⁺PD-1⁺, FOXP3⁺ cell densities and PD-L1 TPS and CPS in FFPE tissue sections in a CLIA-certified laboratory (Figure 1A). Validation studies demonstrated a high concordance between ImmunoProfile and IHC (Supplementary Figure 1). The application of ImmunoProfile to clinical samples revealed minimal run-to-run variability in staining, imaging, and scoring over time (Supplementary Figure 2). We observed minimal variation in results across technicians and pathologists, and < 3% of cases failed for technical reasons (Supplementary Methods). The entire multistep workflow met or exceeded clinical laboratory quality-control requirements and was approved for prospective clinical application.

Patients with high intratumoral immune cells and PD-L1 expression are found across cancer types

In a prospective study, we analyzed biopsy and resection tissue specimens sent for pathology review from all consented cancer patients for 3 years. The final 2,023 case-cohort included 14 major cancer types, biopsies from primary and metastatic sites, and patients with low- and high-stage disease, as expected for a tertiary cancer center (Figure 1B, Supplementary Table 1).

We found that intratumoral immune cell numbers and PD-L1 expression varied substantially across cases and cancer types (Figure 1C, Supplementary Figure 3). When we divided cases by their respective biomarker scores into the cohort’s highest, middle, and lowest tertiles, we found cases with individual scores in the top and bottom tertiles for every cancer type (Supplementary Figure 3). Thus, patients can have “hot” or “cold” tumors irrespective of cancer type, although patients with some cancer types are more likely to have a “hot” tumor microenvironment than others.

Immune biomarker scores split by tertiles stratify patient survival in pan-cancer analyses

We next examined the prognostic significance of the reported immune biomarker scores. In a pan-cancer analysis (n = 2,023, 632 deaths), we divided patients by their respective numbers of intratumoral CD8⁺, PD-1⁺, CD8⁺PD-1⁺, and FOXP3⁺ cells into tertiles (Supplementary Table 2). Those in the top tertiles had longer OS than those in the middle, which, in turn, had longer OS than those in the bottom tertile for each immune cell type (all log-rank p < 0.0001, Figure 2A). Patients whose tumors had PD-L1 combined proportion scores (CPS) in the highest and middle tertiles had longer OS than those patients with tumors in the lowest tertile (p = 0.002, Supplementary Figure 4A). In contrast, patients whose tumors had PD-L1 tumor proportion scores (TPS) in the highest, middle, or lowest tertiles had similar OS (Supplementary Figure 4B). These data reveal that intratumoral CD8⁺, PD-1⁺, CD8⁺PD-1⁺, and FOXP3⁺ cell densities and PD-L1 CPS captured in FFPE tissues by this clinical test predict survival in the pan-cancer setting.

• Download figure
• Open in new tab

Figure 2.

Kaplan-Meier estimates of overall survival (OS) with biomarker values divided into tertiles. The pan-cancer and major cancer sub-types were divided according to their CD8⁺, PD-1⁺, CD8⁺PD-1⁺, and FoxP3⁺ immune cell densities into high (blue), middle (green), and low (red) tertiles for each respective group. The number of patients at risk of death at the indicated time from diagnosis is indicated for A: all patients (n = 2,023), and patients with B: non-small cell lung cancer (NSCLC, n = 489), C: colorectal cancer (CRC, n = 228), and D: esophagogastric cancer (EGD, n = 111), The p-values by the log-rank test are indicated.

Immune biomarker scores split into tertiles predict OS for select major cancer types

We next divided patients with non-small cell lung cancer (NSCLC, n = 489), colorectal cancer (CRC, n = 228), esophagogastric cancer (EGC, n = 111), pancreatic cancer (PAN, n = 104), and bladder cancer (BLAD, n= 159) into high, middle, and low groups based on tertile divisions of their respective biomarker scores (Supplementary Table 2). Those with NSCLC, CRC, and EGC and intratumoral CD8⁺, PD-1⁺, and/ or CD8⁺PD-1⁺ cells in the highest tertiles for each group had significantly better OS than those in the middle or lowest tertiles (all significant log-ranks p < 0.05, Figure 2B-D). Patients with PANC and PD-1⁺ cells in the highest and lowest tertiles had better OS than those with PD-1⁺ cells in the middle tertile (p =0.03, Supplementary Figure 5A), which may reflect an unusual immune response profile or a weaker association between intratumoral immune cells and OS for this cancer type.³⁹ Patients with BLAD did not show differences in OS when cases were split in this manner (all p > 0.05, Supplementary Figure 5B). None of the groups showed significant differences in OS when split by PD-L1 CPS and TPS scores (all p > 0.05).

Given the large number of patients with NSCLC, we further split patients into those receiving immune checkpoint blockade (ICB, anti-PD-1, or anti-PD-L1, n = 114) or not (no-ICB, n = 372). For both groups, we found that those with CD8⁺, PD-1⁺, or CD8⁺PD-1⁺ cells in the highest tertiles had longer OS than those in the lowest tertiles (all p < 0.001, Figure 3A-B). Thus, intratumoral CD8⁺, PD-1⁺, CD8⁺PD-1⁺ and/or FOXP3⁺ cell numbers are associated with OS among patients with NSCLC, CRC, EGC, and PANC by this analysis.

• Download figure
• Open in new tab

Figure 3.

Kaplan-Meier curves of overall survival (OS) according to tertiles of the biomarker distributions. The patients were divided according to their CD8⁺, PD-1⁺, CD8⁺PD-1⁺, and FoxP3⁺ immune cell densities into high (blue), middle (green), and low (red) tertiles for each respective group. The number of patients at risk at the indicated number of years from the time of diagnosis is indicated for A: Patients with NSCLC treated with immune checkpoint blockade (NSCLC (ICB), n = 117), B: Patients with NSCLC treated without immune checkpoint blockade (NSLCL (no-ICB), n = 372). The p-values by the log-rank test are indicated.

Immune biomarker scores split at the median predict OS for select major cancer types

We had fewer than 100 patients and/or 50 deaths for patients with hormone-receptor-positive/ HER2/neu-negative breast cancer (HR⁺/HER2^- BRCA, n = 168), estrogen/progesterone/ HER2/neu negative breast cancer (TN BRCA, n = 55), ovarian cancer (OVCA, n = 90), head and neck squamous cell cancer (HNSCC, n = 88), renal cell carcinoma (RCC, n = 79), endometrial cancer (ENDOM, n = 89), and cutaneous melanoma (MEL, n = 98). Therefore, we divided these patients into “high” and “low” groups based on the median values of their respective biomarker scores. Patients with HR⁺/HER2^- BRCA, TN BRCA, and OVCA and high intratumoral CD8⁺ or CD8⁺PD-1⁺ cell numbers had significantly longer OS than those with low numbers (each significant log-rank p < 0.05, Figure 4A-C). Patients with HNSCC and high CD8⁺PD-1⁺ or FOXP3⁺ cell numbers tended toward more favorable OS (p = 0.08 and 0.06, Figure 4D). Patients with RCC, ENDOM, or cutaneous MEL did not show significant differences in OS by this analysis (all p > 0.05).

• Download figure
• Open in new tab

Figure 4.

Kaplan-Meier estimates of overall survival (OS) for biomarker values divided at the median. The major cancer sub-types were divided according to their CD8⁺, PD-1⁺, CD8⁺PD-1⁺, and FoxP3⁺ immune cell densities into high (blue) and low (red) groups relative to each cancer type’s median values. The number of patients at risk of death at the indicated time from diagnosis is indicated for patients with A: hormone receptor-positive/HER2 negative breast cancer (HR+/HER2^- Breast, n = 168), B: triple-negative breast cancer (TN Breast, n = 55), C: ovarian carcinoma (n = 90), and D: head & neck squamous cell carcinoma (n = 88). The p-values by the log-rank test are indicated.

We next divided patients according to their PD-L1 scores. Those with HR⁺/HER2^- BRCA and high CPS had longer OS than those with low CPS (p = 0.002; Supplementary Figure 6A). In contrast, patients with RCC and high TPS had significantly shorter OS than those with low TPS (p = 0.049; Supplementary Figure 6B). Given the low number of patients (35 or less) with additional cancer types (prostate, hepatobiliary, small cell lung cancer), we did not perform survival analyses for these groups. Thus, CD8⁺ and/or CD8⁺PD-1⁺ cell numbers predict the survival of patients with HR⁺/HER2^- BRCA, TN BRCA, and OVCA. In contrast, PD-L1 TPS predicts the survival of patients with RCC.

Multiple biomarkers predict OS after adjustments for stage and other risk factors

We next examined whether individual biomarkers predicted OS after accounting for major clinical risk factors, including cancer type, biopsy site, AJCC stage, age, gender, smoking, and alcohol use (Table 1; Supplementary Figure 7). Importantly, we selected the risk factors used in each cancer-specific model before calculating the results to avoid overfitting (See Methods).

In the pan-cancer analysis, patients with CD8⁺, PD-1⁺, CD8⁺PD-1⁺, or FOXP3⁺ cells in the top tertiles had longer OS than those in lower tertiles after risk factor adjustments (all p = 0.0001, Table 1). In sub-group analyses, those with NSCLC, with or without ICB, CRC, EGC, BLAD, HR⁺/HER2⁺ BRCA, HNSCC, and OVCA and high numbers of CD8⁺, PD-1⁺, CD8⁺PD-1⁺, and/or FOXP3⁺ cells had significantly longer OS than those with low numbers after risk factor adjustments (all significant p < 0.05, Table 1). Intriguingly, for patients with PANC, the association between intratumoral PD-1⁺ cell numbers and OS was highly significant after risk factor adjustments, including stage (p = 0.007). Patients with TN BRCA and high numbers of CD8⁺PD-1⁺ cells trended towards more prolonged OS after adjusting for stage (p = 0.06, Table 1). Patients with RCC, ENDOM, and cutaneous MEL did not show significant differences in OS by this analysis (Table 1).

In the pan-cancer analysis, patients divided into tertiles by PD-L1 CPS or TPS did not show significant differences in OS after risk adjustments (Supplementary Table 3). However, patients with NSCLC, treated with an immune checkpoint inhibitor (anti-PD-1 or anti-PD-L1), and high PD-L1 CPS (top tertile) or TPS (>1%) had longer OS than those with low PD-L1 scores (p = 0.04, 0.05; Supplementary Table 3). Similarly, patients with CRC and high TPS (> 5%) had longer OS than those with low TPS (p = 0.02, Supplementary Table 3). In contrast, patients with RCC and TPS above the median for the group had worse OS than those below after adjusting for stage (p = 0.04, Supplementary Table 3). We conclude that intratumoral CD8⁺, PD-1⁺, and/or CD8⁺PD-1⁺ cell numbers provide prognostic information not captured by major clinical risk factors for patients with NSCLC, CRC, PANC, EGC, BLAD, HR⁺/HER2⁺ BRCA, HNSCC, and OVCA. PD-L1 scores provide prognostic information not captured by major clinical risk factors for patients with NSCLC treated with ICB and patients with CRC and RCC.

Multivariable biomarker analyses reveal specific immune cell types most closely associated with OS

Numbers of CD8⁺, PD-1⁺, and CD8⁺PD-1⁺ cells were highly correlated across cases (r= 0.8-0.92, Supplementary Figure 8). The numbers of FOXP3⁺ cells and PD-L1 scores were less so (r= 0.26-0.63, Supplementary Figure 8). Therefore, we assessed which individual immune biomarker(s) were most strongly and independently associated with the risk of death using multivariable biomarker modeling (see Methods). In the pan-cancer cohort, patients with high numbers of PD-1⁺ or CD8⁺ cells were independently associated with a reduced risk of death after including clinical risk factor adjustments (Table 2). Those with PD-1⁺ cells in the highest tertile had a 35% reduced risk of death, and those in the middle tertile had a 26% reduced risk compared to those in the lowest tertile (p= 0.0009). We found similar relationships for CD8⁺ cells (p = 0.002, Table 2).

For patients with NSCLC, intratumoral CD8⁺PD-1⁺ cell numbers best predicted the risk of death after accounting for risk factors, including stage and smoking. Those with CD8⁺PD-1⁺ cells in the highest and the middle tertiles each had a 56% reduced risk of death compared to those in the bottom tertile (p = 0.0001, Table 2). For those treated with ICB, intratumoral PD-1⁺ cells best predicted the risk of death; for those not treated with ICB, CD8⁺PD-1⁺ cells best predicted the risk of death (p = 0.003, Table 2).

After risk adjustments, including for MSI status, patients with CRC and intratumoral PD-1+ cells in the highest tertile had an 80% reduced risk of death compared with those in the lowest tertile (p= 0.0006, Table 2). Patients with EGC and CD8⁺ cells in the highest tertile had a 59% reduced risk compared to those in the lowest (p= 0.02). Patients with OVCA and numbers of CD8⁺ cells above the median had a 62% reduced risk compared with those below (p = 0.02, Table 2).

Patients with PANC and BLAD had the most significantly reduced risk of death between those in the highest and middle tertiles for PD-1⁺ and FOXP3⁺ cells, respectively. Those with PANC and intratumoral PD-1⁺ cells in the highest tertile had a 64% reduced risk compared to those in the middle (p = 0.007). Patients with BLAD and FOXP3⁺ cells in the highest tertile had a 68% reduced risk compared to those in the middle (p = 0.03, Table 2). Thus, we find that specific immune biomarkers, most commonly PD-1⁺ and CD8⁺ immune cell numbers, best predict patients’ risk of death independent of currently recognized risk factors, including stage, for major cancer types, including NSCLC, CRC, EGC, OVCA, PANC, and BLAD.

Univariable analyses reveal that specific immune cell populations are most closely associated with OS

We did not attempt multivariable analysis for patients with cancer types having less than 30 deaths. Instead, we described the risk of death with the biomarker most significantly associated with OS after adjusting for clinical risk factors (as shown in Table 1). Patients with HR⁺ /HER2^- BRCA and high CD8⁺ cell numbers had an 80% lower risk of death than those with low numbers (p = 0.002, Table 3). Patients with TN BRCA and high CD8⁺PD-1⁺ cell numbers had a lower risk of death than those with low numbers that approached significance (p = 0.06; Table 3). Patients with HNSCC and high CD8⁺PD-1⁺ cell numbers had a 56% lower risk than those with low numbers (p = 0.05). Finally, patients with RCC and low TPS had a 60% lower risk of death than those with high TPS (p = 0.04, Table 3). These data indicate that specific intratumoral immune markers best predict patients’ risk of death independent of currently recognized risk factors, including stage, for BRCA, HNSCC, and RCC.