hyperPICASSO: adapting the mutual-information-based unmixing technique PICASSO to hyperspectral datasets

Mutual-information (MI)-based unmixing algorithms such as PICASSO (Process of ultra-multiplexed Imaging of biomoleCules viA the unmixing of the Signals of Spectrally Overlapping fluorophores), which iteratively subtract pairs of images to minimize MI, have demonstrated the ability to remove spectral overlap from highly multiplexed fluorescent probes better than reference-based unmixing due to the binding-site variation of fluorophore emission spectra. However, when dealing with hyperspectral datasets, these approaches rely on naïve binning across discrete spectral bands. In doing so, spectral information within those bands is disregarded, resulting in inefficient unmixing. To address this, we have developed the hyperPICASSO algorithm, which generalizes MI-based unmixing to hyperspectral datasets. Our approach is to extract spectral features from annotated images that are used to perform approximate linear unmixing, which is then iteratively improved by pairwise minimization of MI. We find that hyperPICASSO significantly reduces cross talk compared to applying PICASSO-based unmixing to naïve binning of spectral data and compared to linear unmixing using measured spectra. The advantage is particularly evident for features with strongly overlapping emission spectra. We demonstrate that MI-based unmixing can greatly reduce cross talk by utilizing hyperspectral data.