Abstract
Droplet-based profiling techniques such as CITE-seq are often contaminated by technical noise. Current computational denoising methods have serious limitations, including a strong reliance on often-unavailable empty droplets or null controls and insufficient efficiency due to ignoring protein-protein interactions. Here, we introduce scPDA, a probabilistic model that employs a variational autoencoder to achieve high computational efficiency. scPDA eliminates the use of empty droplets and shares information across proteins to increase denoising efficiency. Compared to currently available methods, scPDA substantially improves the efficiency of gating-strategy-based cell-type identification, marking a clear advancement in computational denoising of the protein modality.