Fig. 1: RLSU outperforms linear unmixing and non-negative matrix factorization.

a, Simulated ground-truth data of eight objects (letters of the word SPECTRUM) labelled with different fluorophores. b, Simulated acquired data, spectrally mixed and including Poisson (shot) noise. c, Reconstructed objects using data from b and linear unmixing. Negative values are shown in red. d, Reconstructed objects using data from b and one iteration of RLSU. At this stage, the algorithm has not converged on a solution. e, Reconstructed objects after convergence using data from b and 100 iterations of RLSU. f, Linearly unmixed objects from an experimentally acquired dataset of a U2OS cell co-expressing six different fluorescent protein species. Negative values are shown in red. g, RLSU unmixed objects from the dataset in f. h, Non-negative matrix factorization (NMF)-unmixed objects from the dataset in f. Note that signals have not been correctly assigned to objects (for example, the nuclear signal in the plasma membrane channel). i, Comparison of linear unmixing and RLSU for another real dataset in which mitochondrial signal (magenta arrowheads) bleeds through to the nuclear channel after linear unmixing, but is correctly absent after RLSU. Scale bars, 10 μm.