Bulk tumor RNA can now be split into malignant cell states: 193 states found across 33 cancer types

A bulk RNA tumor sample can now be decomposed into multiple malignant cell states rather than a single generic malignant component. Earlier unsupervised deconvolution methods mainly resolved broad cell types and typically returned one malignant component, leaving heterogeneity inside the malignant compartment uncharacterized. On simulated bulk data from glioblastoma, breast, ovarian, lung adenocarcinoma, and squamous cell carcinoma, CDState ranked first on the overall score combining proportion and expression metrics, with a total score of 3.56, a correlation of 0.54 with true proportions and 0.63 with true expression profiles.

The decomposition needs neither prebuilt reference data nor predefined marker genes. Reference-dependent supervised methods require the malignant states to be present in the reference, which many cancer types lack. On four synthetic mixtures, CDState had lower proportion error than the supervised InstaPrism run with matched single-cell reference in three of four mixtures and higher correlation in two of four.

The method adds two constraints to nonnegative matrix factorization and optimizes them in two annealing-like steps. Classic nonnegative matrix factorization does not constrain component weights, so they cannot be read as cell proportions, and components can remain highly similar. A sum-to-one constraint on per-sample proportions places each sample inside a simplex whose vertices are pure sources; a penalty on pairwise cosine similarity between source vectors pushes transcriptionally similar states apart, and input tumor purity monitors the optimization so the penalty is rolled back once purity error starts to rise.

Applied to 9,669 samples across 33 TCGA cancer types, the method produced 193 malignant cell states and linked their proportions to survival, treatment response, and genetic alterations. Earlier bulk atlas analyses did not systematically characterize the state composition inside the malignant compartment. Malignant components correlated strongly with tumor purity in 27 of 33 datasets; at least one state was significantly associated with worse prognosis in 17 cancer types, 22 states in total, and with better survival in 10 cancer types, 14 states in total.

The method targets tumor cohorts that already have bulk RNA sequencing and tumor purity estimates, and it outputs malignant cell states shared across samples with their proportions rather than single-cell resolution. In the glioblastoma simulated data, highly collinear states merged into one source, for example OPC-like, NPC-like-2, and cycling cells, and MES-like-2 with AC-like. The TCGA analysis covers 33 cancer types and 9,669 samples, identifying 193 malignant cell states.

The method depends on input tumor purity estimates, and output varies with their reliability; the TCGA analysis used ABSOLUTE purity from a different tissue aliquot than the RNA sequencing, which may introduce composition differences. The number of components k is chosen by maximizing correlation between inferred malignant proportions and input purity, which prioritizes separating malignant from nonmalignant compartments but does not guarantee fine resolution of all microenvironment cell types. When the true number of states is unknown, some ground-truth states were not resolved into distinct sources. The candidate genetic drivers come mainly from correlation analyses, which the authors say require experimental validation; treatment-response associations come from limited datasets, which the authors say warrant study in larger independent cohorts. In the CheckMate 009 clear cell renal cell carcinoma data, the heterogeneity index showed a trend toward higher values in patients with clinical benefit but did not reach statistical significance.