Abstract PS3-04-25: Artificial intelligence (AI) based image analysis of PD-L1, TIL, immune signature and ctDNA for prediction of response to neoadjuvant chemotherapy in breast cancer

Background Tumor-infiltrating lymphocytes (TILs) have been recognized as prognostic and predictive biomarkers in breast cancer. However, challenges such as interobserver variability and lack of standardized cut-off values have limited their clinical implementation. Recent advancements in artificial intelligence (AI)-based pathology have shown potential to enhance reproducibility and objectivity in TIL quantification. Concurrently, circulating tumor DNA (ctDNA) has emerged as a minimally invasive biomarker capable of detecting treatment response after neoadjuvant chemotherapy (NAC). Objectives This study aims to: (1) evaluate the predictive value of AI-quantified stromal TILs from H&E-stained slides; and (2) validate the clinical utility of ctDNA-based biomarkers for detection of responses in breast cancer patients after NAC. Methods TILs were assessed from pre-treatment tumor H&E slides using Lunit SCOPE IO, a validated AI-powered WSI analysis tool. Immune phenotypes were classified as inflamed, immune-excluded, or desert. PD-L1 expression was evaluated using PD-L1 IHC 22C3 pharmDx. TIL cut-off was set at 30%, Formalin-fixed paraffin-embedded samples (FFPEs) were obtained from biopsy tissue, and plasma samples were collected at baseline, pre-surgery (after NAC). Forty-seven breast cancer-related genes were analyzed using next-generation sequencing. The diagnostic performance of ctDNA was evaluated, and logistic regression analyses were conducted to assess the impact of clinical and molecular factors on ctDNA status. Parallel plasma samples were collected at predefined intervals for ctDNA extraction and sequencing using a customized Breast NGS panel targeting 157 hotspot mutations across 12 genes. Mutation validation was performed using ddPCR. Preliminary Results A total of 101 patients with early breast cancer were enrolled. The most frequently identified gene was TP53 (FFPE, 66.7%; ctDNA, 46.4%), followed by PIK3CA (FFPE, 36.2%;ctDNA, 17.4%). The triple-negative breast cancer (TNBC) subtype exhibited the strongest association with ctDNA detection (odds ratio [OR] 209.50, p = 0.005) in multivariate analysis. Patients with inflamed TIL phenotype at diagnosis and ctDNA clearance after NAC had higher pathological complete response (pCR) rate (38.5% vs. 11.1%, p = 0.238). Conclusion Our integrated approach using AI-based TIL quantification and ctDNA analysis using NGS and ddPCR provides a feasible and sensitive approach for NAC response in breast cancer patients. These findings support the clinical utility of combining digital pathology and liquid biopsy for treatment monitoring and risk stratification in conjunction with precision oncology. Citation Format: E. Kim, I. Do, S. Chae. Artificial intelligence (AI) based image analysis of PD-L1, TIL, immune signature and ctDNA for prediction of response to neoadjuvant chemotherapy in breast cancer [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-04-25.