Spatial architecture of tumor-infiltrating lymphocytes adds predictive value beyond stromal TIL density for neoadjuvant response in triple-negative breast cancer

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background Stromal tumor-infiltrating lymphocytes (TILs) are established biomarkers in triple-negative breast cancer (TNBC), but conventional density-based assessment does not capture their spatial relationship to tumor nests. We evaluated whether an H&E-derived spatial architecture index (SAI) provides predictive information beyond stromal TIL density for pathologic complete response (pCR) after neoadjuvant therapy. Methods This single-center retrospective cohort included 236 patients with TNBC who had evaluable pretreatment core biopsy slides and definitive surgical response assessment. The primary SAI was the equally weighted mean of standardized close interaction ratio and lymphocyte cluster index, together with reverse-coded standardized lymphocyte–tumor distance and edge enrichment. The primary multivariable analysis included 231 patients with complete clinicopathologic data. Robustness was evaluated using bootstrap optimism correction, alternative SAI constructions, treatment-regimen and propensity-score analyses, repeated nested cross-validation, and a 60-case reproducibility assessment. During internal validation, preprocessing and SAI construction were repeated within each training fold. Results Of 236 patients, 122 (51.7%) achieved pCR. In the main multivariable model, the SAI remained associated with pCR (OR 2.79 per 1 SD, 95% CI 1.79–4.35; P < 0.001), whereas stromal TIL density was not independently significant (OR 0.88 per 10% increase, 95% CI 0.74–1.04; P = 0.123). Bootstrap optimism-corrected AUROCs were 0.698 for the clinical model, 0.705 after the addition of stromal TIL density, 0.764 after the addition of the SAI, and 0.765 after the addition of both. Across six internally validated algorithms using the full feature set, mean AUROCs ranged from 0.730 to 0.752, with only modest between-algorithm differences. The SAI showed excellent interobserver reproducibility (ICC 0.93, 95% CI 0.89–0.95). Its association with pCR persisted in the chemotherapy-only subgroup (OR 2.48, 95% CI 1.52–4.05; P < 0.001), whereas the interaction with pembrolizumab-containing treatment was not significant ( P = 0.222). Conclusion In this exploratory single-center retrospective cohort, the H&E-derived SAI provided internally validated predictive information beyond stromal TIL density for pCR. External multicenter validation, assessment of intersite technical reproducibility, and prospective calibration are required before clinical implementation.