Abstract PD7-09: Feasibility of a machine learning-based peripheral blood immunoprofiling platform to stratify patients with early-stage triple-negative breast cancer (eTNBC) by neoadjuvant therapy response

Abstract Background: Tissue-based biomarkers for predicting breast cancer outcomes are limited by sampling bias and tumor heterogeneity. Blood-based immunoprofiling offers a promising alternative. A machine learning-based platform using multiparameter flow cytometry (MFC) analysis of peripheral blood mononuclear cells (PBMCs) previously identified 5 immunotypes — G1-Naïve (Naïve T cell enriched), G2-Primed (Memory CD4+ T cell enriched), G3-Progressive (DC/NK/Memory CD8+ T cell enriched), G4-Chronic (Effector/Exhausted CD8+ T cell enriched), G5-Suppressive (Myeloid/MDSC enriched), with potential associations to treatment outcomes (Dyikanov et al. Cancer Cell 2024). We aimed to assess whether applying this platform to serial PBMC samples from 2 eTNBC cohorts treated with neoadjuvant therapy (NAT) (1: chemotherapy ChT; 2: ChT+immunotherapy IO) could stratify patients (pts) into response groups. Methods: We identified from the prospective DFCI Multicenter TNBC registry pts with eTNBC, available pathological response information, and PBMCs at both baseline (treatment-naive) and on-NAT (weeks 4-7). PBMCs were stained with 10 custom antibody panels (mean 14 antibodies/panel) and analyzed by MFC. A regression model calculated immune signature scores (ISS; 0: least aligned, 10: most aligned), assigning samples to one of 5 immunotypes based on maximum ISS. Associations with pathologic complete response (pCR, yes/no) were assessed by Fisher’s exact test, Mann-Whitney U test (p.05), and ROC-AUC. Results: 51 pts were evaluated, with clinicopathologic characteristics summarized in the Table. 49 baseline and 46 on-NAT PBMC samples were processed. Overall, immunotypes were not significantly associated with pCR in ChT or ChT+IO groups at any timepoint. ROC-AUC showed modest predictive performance. However, immunotype distribution differed significantly by therapy type at both timepoints (p.05) and was more stable from baseline to on-NAT in ChT (61%) than in ChT+IO (24%). In line with this, in the ChT+IO group, pts experiencing pCR tended to exhibit G1-Naïve on-NAT, regardless of baseline immunotype. Non-pCR pts tended to have G2-Primed at baseline. Consistently, ISS modestly associated with response for G1-Naïve on-NAT (p=.09) and G2-Primed at baseline (p=.02). In both cohorts, CD8+ T-cell subsets were prevalent in pCR pts, while CD4+ subsets were more frequent in non-pCR pts at both timepoints. Baseline Th1/Th2 ratio significantly distinguished pCR from non-pCR (overall p=.009; ChT p=.02; ChT+IO p=.21). Conclusions: Although immunotypes of PBMCs did not reliably predict pCR, their distribution differed by treatment and evolved during NAT, especially in pts on ChT+IO. These findings suggest that immune dynamics may better predict therapeutic benefit and warrant validation in larger studies. Citation Format: C. Corti, A. R. Martin, T. Rahman, A. Patel, A. Rajoo, L. H. Santa Ines, A. M. Parsons, M. F. Goldberg, A. Bolshakova, D. Tumasyan, A. Ryabykh, J. K. Lennerz, N. Ahmad, N. M. Tung, N. Sinclair, M. A. Faggen, S. Sinclair, M. Costantinou, S. Lo, J. L. Meisel, E. Winer, R. Salgado, N. U. Lin, S. S. McAllister, S. M. Tolaney, A. C. Garrido-Castro, E. A. Mittendorf. Feasibility of a machine learning-based peripheral blood immunoprofiling platform to stratify patients with early-stage triple-negative breast cancer (eTNBC) by neoadjuvant therapy response 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 PD7-09.