Deconvolution of heterogeneous bulk tumor samples into distinct cellular populations is an important yet challenging problem, particularly when only partial references are available. A common approach to dealing with this problem is to deconvolve the mixed signals using available references and leverage the remaining signal as a new cell component. However, as indicated in our simulation, such an approach tends to over-estimate the proportions of known cell types and fails to detect novel cell types. Here, we propose PREDE, a partial reference-based deconvolution method using an iterative non-negative matrix factorization algorithm. Our method is verified to be effective in estimating cell proportions and expression profiles of unknown cell types based on simulated datasets at a variety of parameter settings. Applying our method to TCGA tumor samples, we found that proportions of pure cancer cells better indicate different subtypes of tumor samples. We also detected several cell types for each cancer type whose proportions successfully predicted patient survival. Our method makes a significant contribution to deconvolution of heterogeneous tumor samples and could be widely applied to varieties of high throughput bulk data. PREDE is implemented in R and is freely available from GitHub (https://xiaoqizheng.github.io/PREDE).