Starfysh reveals heterogeneous spatial dynamics in the breast tumor microenvironment

Abstract

Spatially-resolved gene expression profiling provides valuable insight into tissue organization and cell-cell crosstalk; however, spatial transcriptomics (ST) lacks single-cell resolution. Current ST analysis methods require single-cell RNA sequencing data as a reference for a rigorous interpretation of cell states and do not utilize associated histology images. Significant sample variation further complicates the integration of ST datasets, which is essential for identifying commonalities across tissues or altered cellular wiring in disease. Here, we present Starfysh, the first comprehensive computational toolbox for joint modeling of ST and histology data, dissection of refined cell states, and systematic integration of multiple ST datasets from complex tissues. Starfysh uses an auxiliary deep generative model that incorporates archetypal analysis and any known cell state markers to avoid the need for a single-cell-resolution reference in characterizing known or novel tissue-specific cell states. Additionally, Starfysh improves the characterization of spatial dynamics in complex tissues by leveraging histology images and enables the comparison of niches as spatial "hubs" across tissues. Integrative analysis of primary estrogen receptor-positive (ER+) breast cancer, triple-negative breast cancer (TNBC), and metaplastic breast cancer (MBC) tumors using Starfysh led to the identification of heterogeneous patient- and disease-specific hubs as well as a shared stromal hub with varying spatial orientation. Our results show the ability to delineate the spatial co-evolution of tumor and immune cell states and their crosstalk underlying intratumoral heterogeneity in TNBC and revealed metabolic reprogramming shaping immunosuppressive hubs in aggressive MBC. Starfysh is publicly available (https://github.com/azizilab/starfysh).

SpottedPy quantifies relationships between spatial transcriptomic hotspots and uncovers new environmental cues of epithelial-mesenchymal plasticity in cancer

“…By adopting our SpottedPy methodology to explore the spatial dynamics of tumour plasticity phenotypes in breast cancer, we have uncovered key differences between tumour regions undergoing EMT and those lacking evidence for this transformation. Our approach illuminates the pronounced spatial correlations of EMT with key cancer hallmarks, notably hypoxia and angiogenesis, in line with findings by He et al 11 in breast cancer spatial transcriptomics detecting these signatures overlapping certain niches. As tumour cells undergo EMT in response to hypoxic stimuli, they are likely to gain a survival advantage in a nutrient-deprived environment and be better equipped to invade and migrate towards regions with better oxygenation, potentially following angiogenic gradients 60,43,46 .…”

“…Macrophages secrete TGF-β, TNF-α, IL-6 and IL-8, which are all well characterised EMT stimuli 64,65 . The relationship has been observed in bulk transcriptomics 66 , in spatial analysis of mouse models of skin carcinoma where depletion of macrophages inhibited EMT progression 31 and in specific niches within breast cancer spatial transcriptomics slides 11 .…”

“…Profiling the cancer tissue spatially allows us to explore the tumour architecture and its heterogeneity in detail, paving the way to deciphering the crosstalk between tumour cells and their surroundings and opening up new therapeutic opportunities 6,7 . Several studies to date have demonstrated the advantages of spatial transcriptomics in delineating major tissue domains with distinct cell composition 8 , cancer hallmarks 9 , immunosuppressive hubs 10,11 , entire tumour ecotypes with divergent clinical outcomes 12 or the impact of specific drugs on inhibiting tumour progression 13 . However, focusing on areas of the tissue that are relevant for a specific biological question and surveying relationships between cell populations at the right scale remains a challenge in these datasets.…”

Benchmarking the translational potential of spatial gene expression prediction from histology

“…Methods that performed joint modelling of H&E (e.g. Cottrazm 43 , starfysh 44 , and xfuse 13 ) were also excluded as we were only investigating the applicability of H&E to ST. In addition, methods predicting or incorporating scRNA-seq (e.g.…”

Artificial intelligence classifies and predicts the outcome of primary CNS and nodal large B-cell lymphomas

“…This gene list has been previously published and it is available in Table S14. We applied the recently described Spatial Transcriptomic Analysis using Reference-Free auxiliarY deep generative modeling and Shared Histology (starfysh) 27 algorithm that implements a joint modeling of transcriptomic measurements and histology images, to infer different cell states employing spatial transcriptomics. Interestingly, using tumor enriched PCNSL samples Figure 4A-E, we identi ed a high degree of intra-tumor heterogeneity, because the 4 molecular groups were present in the same sample, Figure 4F.…”