Abstract GS1-05: Prognostic and predictive associations of manual, digital and AI-derived tumor infiltrating lymphocytes-scoring: A retrospective analysis from the Phase III APHINITY trial

Abstract Background: Stromal tumor-infiltrating lymphocytes (sTILs) are prognostic and predictive biomarkers for HER2-targeted therapy in early-stage HER2-positive breast cancer (BC). Manual sTIL scoring demonstrates high reproducibility but may underrepresent immune infiltration. Digital pathology and artificial intelligence (AI) offer automated sTIL quantification and spatial assessment but require validation against clinical endpoints. Aim: To compare manual, digital (non-AI), and AI-based sTIL quantification methods, including AI-derived spatial metrics in the phase III APHINITY trial. Objectives included interobserver reproducibility, method concordance, prognostic performance for invasive disease-free survival (iDFS) and overall survival, enhancement of prognostic models by AI, and identification of patients benefiting from adjuvant pertuzumab. Methods: Of 4,804 APHINITY trial participants, 4,306 (90%) had evaluable archival H&E whole-slide images for all methods. Manual scoring was performed by an experienced pathologist (RS), with reproducibility assessed on 262 slides independently reviewed by five pathologists using international TIL guidelines. Digital quantification used a standard image-analysis algorithm, and AI-percentage lymphocyte scoring employed Case45’s zero-shot deep-learning biomarker pipeline. Two AI-derived spatial metrics were computed: density of lymphocytes adjacent to cancer nests and quantification of dense lymphoid aggregates (immune hotspots). High-TIL status was defined as ≥75th percentile for percentage scores and ≥50th percentile for spatial metrics. Concordance was measured using intraclass correlation coefficient (ICC) and pairwise agreement. Prognostic and predictive performance were assessed via multivariable Cox regression adjusted for treatment arm, age, chemotherapy regimen, hormone-receptor status, nodal status, tumor size, and grade, with likelihood-ratio tests for incremental prognostic gains. Predictive benefit was evaluated by treatment-by-TIL interaction hazard ratios (HR) for iDFS in high-TIL groups. Results: Manual scoring reproducibility was excellent (ICC 0.87). Continuous sTIL quantification showed modest to moderate concordance across methods (ICC 0.37-0.62), with digital and AI consistently scoring lower than manual. High/low TIL classification achieved 80% overall agreement. Every 10% increment in sTIL was associated with reduced recurrence risk (HR: manual 0.93, digital 0.92, AI 0.87; all p < 0.001). AI spatial immune hotspots outperformed percentage metrics (HR = 0.41; p < 0.001) and, when combined with manual scores, provided the greatest additional discrimination (p < 0.001). In high-TIL patients, pertuzumab reduced iDFS events by 64% (manual HR = 0.36; p int = 0.003), 52% (digital HR = 0.48; p int = 0.025), and 54% (AI HR = 0.46; p int = 0.01). Manual scoring alone identified 562/2,573 (22%) node-positive patients as high-TIL and likely pertuzumab-responsive, whereas AI-percentage lymphocyte identified 625 (24%) thus contributing to further detect 253 node-positive patients who would benefit from addition of pertuzumab (a 10% larger group). AI spatial metrics were not predictive. Conclusions: Despite modest concordance, manual, digital and AI-derived sTIL assessments independently demonstrated prognostic and predictive value for identifying HER2-positive BC patients benefiting from adjuvant pertuzumab. AI-driven sTIL quantification matches and slightly improves prognostic accuracy, importantly both AI and manual sTIL can identify a larger group of pertuzumab-responsive patients. Integrating AI-derived quantitative and spatial metrics into multiparameter models can further individualize HER2-targeting therapy. Citation Format: R. Salgado, L. E. Lara Gonzalez, F. Giudici, F. Rojo, L. Comerma, S. Wienert, J. Palacios, Z. Kos, S. L. De Haas, A. Rodriguez Lescure, G. Viale, Y. Zheng, D. Gao, A. Kiermaier, F. Andre, S. Loibl, M. J. PIccart, R. Gelber, D. Cameron, I. E. Krop, P. Savas, T. O. Nielsen, C. Denkert, S. Michiels, S. Loi, APHINITY Steering Committee and Investigators, The International Immuno-Oncology Biomarker Working Group..Prognostic and predictive associations of manual, digital and AI-derived tumor infiltrating lymphocytes-scoring: A retrospective analysis from the Phase III APHINITY trial [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 GS1-05.