Single-cell data often contain dropout noise and nonlinear trajectories. This chapter presents a kernel non-negative matrix factorization (KNMF) framework that first builds a cell-pair differentiability correlation kernel, and then decomposes it into low-rank metagene and cell-loading matrices while preserving non-negativity. A diffusion-map-guided variance criterion automatically selects the cluster number. Tests on neuronal, pluripotent and PBMC datasets show superior accuracy and stability over classical non-negative matrix factorization frameworks, k-means and spectral alternatives, enabling robust dissection of cellular heterogeneity.