Fig. 1

Schematic of our multi-study NMF framework. A Under the discovery-only model, we take the counts matrices for any number of studies as input, and learn signatures, exposures, and indicators of which signatures belong to each study. In this illustration, two studies are decomposed into a total of three signatures; the top study is found to contain the first two signatures, whereas the bottom study is found to contain the first and third signatures. This sharing pattern is shown by the signature indicator matrix. B Under the recovery-discovery model, we expand the NMF decomposition to both recover previously known signatures, encoded via informative priors, and to discover new signatures. Here, we show an example decomposition for one study, where one signature (SBS 2) is recovered and one signature is discovered. C In both models, we can incorporate covariates, such as gender and age as illustrated here, to learn relationships with the exposures