INTRODUCTION: The estrogen receptor (ER) is routinely assessed by immunohistochemistry (IHC) in breast cancer to stratify patients into therapeutic and prognostic groups. Pathology laboratories are burdened by an increased number of biopsies, and costly and resource-demanding molecular pathology analyses. Automatic, artificial intelligence-based prediction of biological properties from hematoxylin and eosin (HE)-stained slides could increase efficiency and potentially reduce costs at laboratories. The aim of this study was to develop a model for prediction of ER status from HE-stained tissue microarrays (TMAs). Our methodology can be used as proof-of-concept for the prediction of more complex and costly molecular analyses in cancer. METHODS: In this study, TMAs from more than 2,000 Norwegian breast cancer patients were used to train and predict ER status using the clustering-constrained attention multiple-instance learning (CLAM) framework. Two patch sizes were evaluated, multi-branch and single-branch CLAM configurations were compared, and a comprehensive hyperparameter search with more than 16 000 experiments was performed. The models were evaluated on internal and external test sets. RESULTS: On the internal test set, the proposed model achieved a micro accuracy, a macro accuracy, and an area under the curve of 0.91, 0.86, and 0.95, respectively. The corresponding results on the external test set were 0.93, 0.76, and 0.91, respectively. Using larger patch sizes resulted in significantly better classification performance, while no significant differences were observed when changing CLAM configurations.