The tumor microenvironment (TME) influences treatment response in breast cancer, yet identifying spatial patterns predictive of treatment response remains a challenge. Here we applied cyclic immunofluorescence proteomics to phenotype approximately 5.5×105 single cells from pre-treatment tissue biopsies of 13 ER-positive breast cancer patients. For each cancer cell, a radial distribution-based neighborhood was constructed by calculating the distribution of identified phenotypes up to a radius of 200 µm, resulting in a total of 2.0×105 neighborhoods. To identify responder-associated neighborhood patterns, an ensemble of neural networks combining autoencoding and multiple-instance learning was trained on sample-level response labels to neoadjuvant chemotherapy with bevacizumab. The resulting model distinguished responders from non-responders in cross-validation, ranked held-out responders above non-responders, and revealed spatially heterogeneous responder-like areas within tumors. These findings demonstrate that weakly supervised learning of spatial proteomic data can identify clinically relevant TME structures associated with treatment response.
Publication Date: 2026-05-22