A unique malignant cell type per patient tumor is encoded in each cancer cell transcriptome.

Deciphering shared features between patients through unsupervised analyses of tumor single-cell transcriptomes is hindered by the predominant clustering of malignant cells based on the patients tumor of origin. In contrast, cancer-associated non-malignant cells within tumors cluster according to their somatic cell type (e.g. macrophage, fibroblast, astrocyte), regardless of the individual patient. We investigated the origin of this contrasting clustering behavior using computational analyses and novel data sampling techniques across 160 tumors representing 9 distinct cancer types. This led to the unexpected disclosing of tumor-specific malignant cell types. We demonstrate that tumor-driven malignant cell clustering is independent from technical or computational biases, nor is it reducible to defined gene sets with tumor-specific expression across all examined cancers. On the opposite, we unveil redundant information dispersed across the transcriptome that encodes a unique identity shared by the malignant cells within each tumor. Additionally, we found a similar genomic encoding for the identity of normal cell types. Finally, we demonstrate that malignant cell identities are maintained across space and over time, as are the identities of normal somatic cells types. These findings suggest the establishment of a distinct type of malignant cells within each tumor, robustly and diffusively encoded across the entire transcribed genome.