This study proposes a cell deconvolution framework based on mixtures of non-negative Beta regression models that captures latent heterogeneity in DNA methylation profiles and significantly improves cell proportion estimation compared with standard deconvolution approaches.