Immunoprofiling at an Institutional Scale Reveals That High Numbers of Intratumoral CD8+ and PD-1+ Cells Predict Superior Patient Survival Across Major Cancer Types Independent of Major Risk Factors

Abstract

Purpose: Retrospective studies have found associations between the number of intratumoral immune cells and patient outcomes for specific cancers treated with targeted therapies. However, the clinical value of routinely quantifying intratumoral immune biomarkers using a digital pathology platform in the pan-cancer setting within an active clinical laboratory has not been established.

Methods: We developed ImmunoProfile, a daily clinical workflow that integrates automated multiplex immunofluorescence tissue staining, digital slide imaging, and machine learning-assisted scoring to quantify intratumoral CD8⁺, PD-1⁺, CD8⁺PD-1⁺, and FOXP3⁺ immune cells and PD-L1 expression in formalin-fixed, paraffin-embedded tissue samples in a standardized and reproducible manner. We prospectively applied ImmunoProfile to biopsies collected from 2,023 unselected patients with cancer over a 3-year period in the clinical laboratory and correlated the results with patient survival.

Results: In the pan-cancer cohort, patients with high numbers of intratumoral CD8⁺ or PD-1⁺ cells in had significantly lower risks of death compared with those with low numbers (CD8⁺: high v low hazard ratio [HR], 0.62 [95% CI, 0.48 to 0.81], Wald P = .002; PD-1⁺: high v low HR, 0.65 [95% CI, 0.51 to 0.83]; P = .0009) after adjusting for risk factors, including cancer type. In subset analyses, patients with high numbers of intratumoral CD8⁺, PD-1⁺, and/or CD8⁺PD-1⁺ cells showed lower risks of death from non-small cell lung, colorectal, breast, esophagogastric, head and neck, pancreatic, and ovarian cancers after considering clinical risk factors, including American Joint Committee on Cancer stage, and despite varying therapies (all P < .05).

Conclusion: Routinely quantifying intratumoral CD8⁺ and PD-1⁺ cells with a clinically validated digital pathology platform predicts patient survival across major cancer types, independent of clinical stage and despite diverse treatment regimens.

FIG 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 ROIs 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 v metastatic) and patients' clinical stage. Further details are provided in the Data Supplement. (C) Violin plots showing the distributions of intratumoral immune cell densities (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 75th and 25th percentiles for the pan-cancer cohort. The black lines represent the mean value, the red boxes represent the 75th and 25th 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. CLIA, clinical laboratory improvement amendments; H&E, hematoxylin and eosin; NSCLC, non–small cell lung carcinoma; RCC, renal cell carcinoma; ROIs, regions of interest.

FIG 2.

Kaplan-Meier estimates of OS with biomarker values divided into tertiles. The pan-cancer and major cancer subtypes were divided into high (blue), middle (green), and low (red) tertiles on the basis of their CD8⁺, PD-1⁺, CD8⁺PD-1⁺, and FoxP3⁺ immune cell densities, respectively, 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) NSCLC (n = 489), (C) CRC (n = 228), and (D) esophagogastric carcinoma (EGD, n = 111). The P values by the log-rank test are indicated. CRC, colorectal carcinoma; NSCLC, non–small cell lung carcinoma; OS, overall survival.

FIG 3.

Kaplan-Meier estimates of OS for biomarker values divided at the median. The major cancer subtypes 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– breast carcinoma (n = 168), (B) TN breast carcinoma (n = 55), (C) ovarian carcinoma (n = 90), and (D) head and neck squamous cell carcinoma (n = 88). The P values by the log-rank test are indicated. HER2– Breast, human epidermal growth factor receptor 2–negative breast carcinoma; OS, overall survival; TN Breast, triple-negative breast carcinoma.

FIG 4.

Risk of death in multivariable biomarker Cox models with clinical risk factor adjustments for cancers with high case numbers. The clinical risk factors considered in the multivariable modeling for each group are listed in Table 1. BLAD, bladder carcinoma; CRC, colorectal carcinoma; cut. MEL, cutaneous melanoma; EGC, esophagogastric carcinoma; ICB, treatment with immune checkpoint blockade; no-ICB, no treatment with immune checkpoint blockade; NSCLC, non–small cell lung cancer; OS, overall survival; OVCA, ovarian carcinoma; PANC, pancreatic carcinoma.

FIG 5.

Risk of death in univariable biomarker Cox models with clinical risk factor adjustments for cancers with low case numbers. The clinical risk factors considered in the univariate modeling for each group are listed in Table 1. HNSCC, head and neck squamous cell carcinoma; OS, overall survival; RCC, renal cell carcinoma; TN BRCA, triple-negative breast carcinoma; TPS, tumor proportion score.