Lens-free digital in-line holography (LDIH) produces cellular diffraction patterns (holograms) with a large field of view that lens-based microscopes cannot offer. It is a promising diagnostic tool allowing high-throughput cellular analysis. Holograms are, however, too complicated to discern by the human eye, and time-consuming computation is required to recover object images. To directly analyze holograms from LDIH, we developed a novel deep learning architecture called a holographical deep learning network (HoloNet) for cellular phenotyping. The HoloNet extracts large features from diffraction patterns, integrates them with small features from convolutional layers, and outperforms other state-of-the-art deep learning methods for the classification of breast cancer cells as well as the prediction of the intensities of breast cancer markers, ER/PR and HER2 from raw holograms. Moreover, to identify previously unknown subclusters of breast cancer cells, we developed the feature-fusion HoloNet model to integrate diffraction features related to breast cancer cell types and marker intensities of ER/PR and HER2. This hologram embedding allowed us to identify the subclusters within the known breast cancer cells. Some of the subclusters identified in our study have phenotypes shared by multiple breast cancer cell types since they are located near the class boundaries in the feature space. Identifying such rare and subtle phenotypes of breast cancer cells will enable us to perform detailed analyses of the heterogeneity of cell phenotypes for precise breast cancer diagnosis.