Decoding multicellular niche formation in the tumour microenvironment from nonspatial single-cell expression data

Accurate inference of granular cell states that co-occur within the tumour microenvironment (TME) is central to defining pro- and anti-tumour environments. However, to reliably identify recurrent coexisting cell populations it is fundamental to analyze datasets encompassing a substantial number of tumour samples with a resolution sufficient to capture granular cell states. Here, we leverage eight scRNA-seq datasets of pancreatic ductal adenocarcinoma (PDAC) in a unique discovery-validation setup and find reproducible cell states, gene programs, and cellular niches that are predictive of specific clinical outcomes. Across tumours, we show highly consistent co-occurrence of cell states within and between lineages, including those reflecting known and de novo cellular interactions alongside the formation of multi-cellular clusters such as tertiary lymphoid structures. In addition, we develop a novel probabilistic model to quantify multi-cellular communities directly from atlas-scale scRNA-seq datasets. This model identified cellular niches predictive of clinical outcomes including communities associated with response to therapy and with specific KRAS mutations. Together, this work lays the foundation for inferring reproducible multicellular niches directly from large nonspatial scRNA-seq atlases and linking their presence in individual patients to prognosis and therapy response.