Abstract Despite advances in molecular profiling, therapeutic development has been hindered by the inability to identify and target tumour-specific mechanisms without consequence to healthy tissue. Correspondingly, a computational framework capable of accurately distinguishing tumour from non-tumour cells has yet to be developed and cell annotation algorithms are unable to assign integrated genomic and transcriptional profiles to single cells on a cell-by-cell basis. To address these barriers, we developed the Single Cell Rule Association Mining (SCRAM) tool that integrates RNA-inferred genomic alterations with co-occurring cell type signatures for individual cells. Applying SCRAM to glioma, we identified tumour cell trajectories recapitulate temporally-restricted developmental paradigms and feature unique co-occurring identities. Specifically, we validated two previously unreported tumour cell populations with immune and neuronal signatures as hallmarks of human glioma subtypes. In vivo modeling revealed a rare immune-like tumour cell population resembling antigen presenting cells can direct CD8+ T cell responses. In parallel, Patch sequencing studies in human tumours confirmed that neuronal-like glioma cells fire action potentials and represent 40% of IDH1 mutant tumor cells. These studies identified new glioma cell types with functional properties similar to their non-tumour analogues and demonstrate the ability of SCRAM to identify these cell types in unprecedented detail.

Identification of functional immune and neuronal tumour cells in glioma