AI-based assessment of tumor-infiltrating lymphocytes (TILs) in patients with triple-negative breast cancer (TNBC) receiving neoadjuvant chemotherapy versus chemo-immunotherapy.

Abstract

e12583 Background: Most patients with early-stage TNBC receive neoadjuvant chemotherapy and immunotherapy, yet biomarkers to guide treatment selection remain limited. Building on prior manual TIL analyses, we evaluated whether AI-derived TIL and immune spatial metrics are associated with response and survival, and whether these associations differ by treatment regimen. Methods: From the prospective DFCI Multicenter TNBC Registry, we identified patients with stage I–III TNBC who received neoadjuvant therapy. Pretreatment H&E–stained slides were analyzed for stromal TILs using both manual assessment and a zero-shot AI pipeline to quantify multiple immune and spatial features, including AI TIL score, immune hotspots, admixed immune infiltration, and immune habitat. Associations with pathologic complete response (pCR) and event-free survival (EFS) were evaluated using modified Poisson regression and Cox proportional hazards models, respectively. Multiple testing was addressed using the Benjamini-Hochberg false discovery rate method (q-values reported). Results: Baseline H&E slides were available from 105 patients who received neoadjuvant therapy (chemotherapy, n=40, (38%); chemo-immunotherapy, n=65, (62%) and included in this analysis. 12 patients (11%) presented with stage I, 44 (42%) stage II, and 49 (47%) stage III disease. 45 (43%) had lymph node involvement, 89 (85%) of patients received the KN522 regimen, and (15%) received a platinum and taxane-containing regimen. Overall, 61 (58%) patients experienced a pCR: 22 (55%) in the chemotherapy-only cohort and 39 (60%) with chemo-immunotherapy. Higher AI-derived TIL metrics were associated with increased likelihood of experiencing pCR, including AI TILs (RR per standard deviation (SD) 1.22, 95% CI 1.12-1.33, p<0.001, q<0.001), admixed immune infiltration (RR per SD 1.14, 95% CI 0.99-1.31, p=0.0625, q=0.120), immune hotspot (RR per SD 1.14, 95% CI 1.00-1.29, p=0.049, q=0.120), and immune habitat (RR per SD 1.15, 95% CI 0.99-1.33, p=0.0638, q=0.120). Associations were observed in the chemotherapy-only and chemo-immunotherapy cohorts combined, without evidence of a statistically significant interaction between AI-TIL metrics and neoadjuvant regimen. With a median follow-up of 2.6 years, the 3-year EFS was 85.3% in the overall cohort. Higher AI-TILs and immune habitat were associated with improved EFS (q<0.10), though event numbers were limited. Conclusions: These findings are consistent with prior manual TIL analyses and demonstrate that AI-TIL metrics are prognostic in early TNBC, though not predictive of benefit from adding pembrolizumab to neoadjuvant chemotherapy. Additional studies with longer follow-up are needed to validate these findings and further explore the potential role of AI-derived TIL metrics in early-stage TNBC.