Purpose: The different prognostic values of tumor-infiltrating neutrophils (TIN) in different tissue compartments are unknown. In this study, we investigated their different prognostic roles and the underlying mechanism. Experimental Design: We evaluated CD66b þ neutrophils in primary tumors from 341 patients with breast cancer from Sun Yat-sen Memorial Hospital by IHC. The association between stromal and parenchymal neutrophil counts and clinical outcomes was assessed in a training set (170 samples), validated in an internal validation set (171 samples), and further confirmed in an external validation set (105 samples). In addition, we isolated TINs from clinical samples and screened the cytokine profile by antibody microarray. The interaction between neutrophils and tumor cells was investigated in transwell and 3D Matrigel coculture systems. The therapeutic potential of indicated cytokines was evaluated in tumor-bearing immunocompetent mice. Results: We observed that the neutrophils in tumor parenchyma, rather than those in stroma, were an independent poor prognostic factor in the training [HR = 5.00, 95% confidence interval (CI): 2.88-8.68, P < 0.001], internal validation (HR ¼ 3.56, 95% CI: 2.07-6.14, P < 0.001), and external validation set (HR ¼ 5.07, 95% CI: 2.27-11.33, P < 0.001). The mechanistic study revealed that neutrophils induced breast cancer epithelialmesenchymal transition (EMT) via tissue inhibitor of matrix metalloprotease (TIMP-1). Reciprocally, breast cancer cells undergoing EMT enhanced neutrophils' TIMP-1 secretion by CD90 in a cell-contact manner. In vivo, TIMP-1 neutralization or CD90 blockade significantly reduced metastasis. More importantly, TIMP-1 and CD90 were positively correlated in breast cancer (r 2 ¼ 0.6079; P < 0.001) and associated with poor prognosis of patients. Conclusions: Our findings unravel a location-dictated interaction between tumor cells and neutrophils and provide a rationale for new antimetastasis treatments.