Article Text

Download PDFPDF

1308 Artificial intelligence (AI)-powered immune phenotyping based on programmed death ligand 1 (PD-L1) immunohistochemistry (IHC) in triple negative breast cancer (TNBC)
  1. Sangwon Shin,
  2. Gahee Park,
  3. Soo Ick Cho,
  4. Taebum Lee,
  5. Juneyoung Ro,
  6. Seulki Kim,
  7. Seunghwan Shin,
  8. Aaron Valero,
  9. Seonwook Park,
  10. Biagio Brattoli,
  11. Jeongun Ryu,
  12. Changho Ahn,
  13. Siraj Ali and
  14. Chan-Young Ock
  1. Lunit Inc., Seoul, Republic of Korea
  • Journal for ImmunoTherapy of Cancer (JITC) preprint. The copyright holder for this preprint are the authors/funders, who have granted JITC permission to display the preprint. All rights reserved. No reuse allowed without permission.


Background The tumor microenvironment has recently become important in cancer and immune phenotype (IP) has been proposed as a way to assess it. A variety of methods to assess IP in tumor tissue were proposed, mostly from histology slides of IHC including CD3 and CD8 or hematoxylin and eosin (H&E). However, methodology based on spatial tumor-infiltrating lymphocyte (TIL) analysis in PD-L1 IHC has been rarely investigated. Here, we performed AI-based IP classification in TNBC using PD-L1 22C3 IHC whole slide images (WSIs) as well as PD-L1 combined positive score (CPS) and its positivity.

Methods We employed Lunit SCOPE IO, an AI-powered H&E analyzer for spatial TIL analysis, identifying and quantifying TIL within cancer or stromal areas in H&E slides. For PD-L1 IHC analysis, Lunit SCOPE PD-L1 CPS, an AI-powered PD-L1 CPS analyzer was used. This model detects and quantifies PD-L1 status in tumor and immune cells and was developed with 3.35x105 tumor cells and 3.45x105 immune cells from PD-L1 IHC-stained WSI of breast cancer. To validate, 180 pairs of PD-L1 IHC and H&E WSIs from a cohort of TNBC were analyzed. The IPs were classified as inflamed (high TIL density in the cancer area, IIP) or non-inflamed (non-IIP) in H&E and PD-L1 IHC WSIs, using a standardized 0.5x0.5mm2 grid for both and a lymphocyte cutoff of 130/mm2 for AI-based PD-L1 IHC analysis.

Results IPs were classified as inflamed in 69 cases (38.3%) in PD-L1 and 40 cases (22.2%) in H&E. The agreement in IPs between two models was 73.9% (table 1). The median PD-L1 CPS was 10 (interquartile range 2 - 25). AI-based PD-L1 IHC analysis revealed significant differences in median CPS levels between IIP and non-IIP, with values of 32.5 (15.6 - 77.2) and 2.9 (0.6 - 8.6), respectively (p<0.001). Moreover, median values of PD-L1 positivity in IIP were significantly higher than in the non-IIP across cell types: tumor cells (14.9% vs. 1.4%, p<0.001), lymphocytes (36.2% vs. 14.1%, p<0.001), and macrophages (14.9% vs. 9.7%, p=0.001).

Conclusions The IP determined by the AI-powered PD-L1 IHC analyzer showed a high concordance rate with the IP determined by the AI-powered H&E analyzer. Moreover, high PD-L1 expression of each respective cell type of tumor, lymphocyte, and macrophage was observed in the IIP. Consistent with previous knowledge of the IP, our PD-L1 results support use of immune-oncology approaches in this phenotype.

Abstract 1308 Table 1

Analysis of IPs by two AI-powered analyzers

This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.