Immune infiltration features relative to cancer cell clumps as predictors of survival in triple-negative breast cancer.

Abstract

1134 Background: Triple-negative breast cancer (TNBC) is an aggressive subtype with poor prognosis. While immune infiltration and abundance have been established as prognostic indicators, we examine how immune spatial organization and tumor geometry jointly influence survival. Methods: Whole-slide H&E images from 125 TNBC TCGA cases were analyzed using pathology instance learning models to generate tile-based predictions, which were reconstructed into polygon annotations for tumor epithelium, tumor-infiltrating lymphocytes (TILs), and tertiary lymphoid structures (TLS). From these annotations, we extracted novel spatial features describing abundance, fragmentation, dispersion, shape, and boundary properties of immune and tumor regions, as well as immune–tumor distances, overlap gradients, and boundary engagement. Associations with disease-specific survival (DSS) were assessed using univariate Cox models, and significant features were incorporated into a multivariate Cox model. Kaplan–Meier and correlation analyses were performed for dichotomized features. Results: Among all spatial features analyzed, global immune dispersion emerged as a strong indicator of survival. This metric, quantified by the radius of gyration of immune islands, was significantly associated with worse disease-specific survival (DSS) (HR per SD = 1.90; p = 0.03) and demonstrated clear Kaplan–Meier separation (log-rank p = 0.028). Increased tumor boundary complexity, measured by edge density normalized to tissue area, was also associated with poorer DSS (HR per SD = 1.77; p = 0.007). Kaplan–Meier analysis for edge density showed only modest separation (log-rank p = 0.071), consistent with a gradual, continuous risk effect rather than a discrete threshold. Similarly, greater intratumoral immune area correlated with adverse DSS (HR per SD = 1.26; p = 0.020), supporting the interpretation of intratumoral immune accumulation as a marker of ineffective or dysfunctional immune infiltration rather than protective immunity. In a multivariable Cox model, immune dispersion, tumor boundary complexity, and intratumoral immune area retained concordant effect directions and jointly achieved strong discrimination (C-index ≈ 0.74). Correlation analysis confirmed that the dominant spatial features were non-redundant, with minimal multicollinearity (all variance inflation factors < 1.2). Conclusions: In TNBC, survival is not solely dependent on immune abundance and proximity to the tumor region, but on the immune spatial dispersion, aggressive tumor boundaries, and ineffective immune accumulation in the intratumoral space. Novel spatial biomarkers examined here may refine risk stratification beyond conventional metrics.