AI tumor analysis reveals why some breast cancers behave differently than others

Abstract Background: Approximately15% of invasive breast cancers are Invasive Lobular Cancer (ILC) with differingclinical outcomes than non-lobular breast cancer (NLC). Digital pathologyenables high-throughput biomarkers and morphology analysis. The TME,particularly TILs is a robust prognostic factor in localized breast cancer.Here, we detail the prognostic importance of ductal vs lobular histology,manually and with AI, combined with manual and novel AI-derived TME metrics inearly (<5y) and late (>5y) recurrence with 15-years of follow-up inTAILORx. Methods: H&E whole slideimages (WSI) from 8,422 patients were reviewed by 16 breast pathologists who completedcentral review for grade, TILs (www.tilsinbreastcancer.org) and histology (WHOclassification). The same H&E dataset was evaluated with an AI-based CDH1-classifier(Paige.AI). In addition, a zero-shot AI model (Case45) generated a panel of TMEbiomarkers (TIL abundance, spatial TIL levels with proximity to cancer cellsand degree of cancer cell-fibroblast contact) combined into a single TME-riskscore. Clinical outcomes were evaluated using multivariable Cox models adjustedfor age, tumor size, Oncotype DX 21-gene recurrence score (RS; Exact Sciences),adjuvant therapy (endocrine vs chemo-endocrine), and centrally determined grade,histology (ILC vs NLC), and manual TILs. Similar analysis were performed usingthe CDH1-classifier rather than central histology, and TME risk score. Results: Centralpathology review revealed ILC in 11.9% of cases, and NLC in 88.1% (77.7% ductal,10.4% non-ductal). Concordances ratesbetween the pathologists were 0.837 and 0.494 (Fleiss’ Kappa) for histology andgrade, respectively, with ICC of 0.968 for TILs. Both centralized pathologyreview (11.9%) and the CDH1-AI classifier (11.2%) demonstrated that ILC has consistentlyhigher risk of recurrence than NLC between years 5-15, but not before; there was a 4.9% overall survival (OS)-difference between ILC vs NLC at 15years. For manual TILs, estimated hazard ratio (HR) for a 10-point differencein TILs was 1.06 (95% CI 1.00, 1.13 p=0.04) for distance recurrence freeinterval (DRFI). AI-derived TME-risk stratified DRFI from 95.7% to 90.9% at 10years, and from 92.1% to 86.9% at 15 years (HR per standard deviation (SD)1.27, 95% CI 1.15-1.40, p<0.0001). This association remained significantafter adjustment for clinicopathologic factors including 21-gene RS (HR per SD1.14, 95% CI 1.04-1.25, p=0.005). Conclusion: ILC(identified by manual review or CDH1-AI classifier) is associated with higherlate recurrence risk and death than NLC at 15 years after diagnosis in TAILORx, the majority of whom received a 5-yearcourse of adjuvant endocrine therapy. Furthermore, both manual TILs andAI-derived TME analysis provide independent risk stratification in addition to 21-geneRS in HR+/HER2− node-negative breast cancer. These findings haveimplications for considering up to a 10-year course of adjuvant ET in womenwith ER+, HER2−, node-negative ILC, even when there is a low 21-gene RS. Citation Format: R. Salgado, R. Gray, G. Broeckx, C. Desmedt, A. Li, G. Van den Eynden, Z. Kos, B. Acs, T. Tramm, E. Stovgaard, C. Focke, L. Comerma Blesa, A. Hida, M. Lacroix-Triki, E. Provenzano, F. Pareja, S. Maley, N. Villena, H. Montgomery, E. Li-Ning-Tapia, A. Lazar, S. Badve, S. Loi, J. Sparano, The International Immuno-Oncology Biomarker Working Group, TAILORx Investigators. Clinical outcomes of invasive lobular carcinoma (ILC) versus non-lobular breast cancer (NLC) assessed by expert pathologists, an artificial intelligence (AI) CDH1 classifier, and AI-derived tumor microenvironment (TME) biomarkers in TAILORx [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 RF3-01.