Artificial intelligence (AI) based spatial assessment of tumor-infiltrating lymphocytes (TIL) and pathologic complete response in early HER2+ breast cancer (BC): Secondary analysis of NSABP B-41.

Abstract

551 Background: Manual quantitative assessment of stromal TILs has shown promise as a biomarker in HER2+ BC. We present an AI-powered single-cell TIL assessment. Methods: Manual TIL assessment was completed per guidelines. Zero-shot, AI-powered pipeline (Case45) was used to analyze tumor microenvironment (TME) from H&E slides, focusing on TILs and their spatial interplay with cancer cells. The algorithm identified all cells, deriving three metrics: pct_lymphocyte (lymphocytes/total cells), AI_TIL (adjacent-tumor lymphocyte to stromal cell ratio), hotspot_immune (normalized fraction of immune cell aggregates in relation to cancer/tissue). Spearman correlation coefficients evaluated correlations; logistic regression models assessed the relationship between TIL measurements and pCR, with and without gene expression adjustments. AUC assessed predictive performance, and univariate Cox models examined TILs' association with event-free survival (EFS). Results: Our analyses included tumors of 262 patients with early-stage HER2+ BC, 67% estrogen receptor (ER) positive, 51% positive lymph nodes. Poor histologic grade (p<0.001), non-luminal (p=0.006), and ER- tumors (p=0.003) were associated with higher manual TILs. Manual TILs were moderately associated with pct_lymphocyte (r= 0.34) and AI_TIL (r= 0.43). Perthe table, manual TILs were positively associated with pCR, the association was numerically stronger in ER- tumors (Interaction p=0.38). pct_lymphocyte and AI_TIL were positively associated with pCR, regardless of ER status. hotspot_immune was strongly associated with pCR (OR=1.26 for all, 1.29 in ER-, 1.22 in ER+, p=<0.001). TILs and ESR1 and ERBB2 provided complementary prognostic utility in pCR in trastuzumab-treated patients (AUC: 0.699-0.757). Among all subjects, there was no association between manual TILs and EFS (p=0.2); there was a marginal association between AI_TIL and EFS (p=0.06). Conclusions: The spatial characterization of TILs using an AI-powered tool shows promise as a prognostic biomarker in both HER2+/ER+ and HER2+/ER- BC, manual TIL assessment is prognostic in HER2+/ER- BC. The assessment of immune aggregates appears to be highly predictive of pCR. Further validation through prospective-retrospective studies, focused on the spatial immune heterogeneity in the TME, is required before integrating these biomarkers into routine clinical practice. Clinical trial information: NCT00486668 . TILs measurements and pCR. Variable (continuous) Cohort OR (95% CI) p-value Manual TILs %(10-unit inc.) All 1.13 (1.04, 1.23) 0.004 ER- 1.16 (1.02, 1.31) 0.02 ER+ 1.07 (0.95, 1.21) 0.27 Percentage of Lymphocyte (10-unit inc.) All 2.00 (1.30, 3.07) 0.002 ER- 1.75 (0.95, 3.21) 0.07 ER+ 1.93 (1.02, 3.62) 0.04 AI TILs(one-tenth inc.) All 1.22 (1.06, 1.40) 0.005 ER- 1.19 (0.98, 1.44) 0.09 ER+ 1.19 (0.97, 1.46) 0.10