RECOMBINE is a computational framework that determines recurrent composite markers of cell types and states. (A) RECOMBINE workflow. (B) Relationship of sparse hierarchical clustering algorithms. (C) The SSL penalty function. The SSL penalty as a function of w with λ₁ = 0.001 and λ₀ ∈ {1, 10, 50}, where w is a one-dimensional (left panel) or two-dimensional variable (other panels). The SSL penalty converges to the LASSO penalty with decreasing λ₀ and to the L0-norm penalty with increasing λ₀. Therefore, the SSL penalty forms a data-adjustable bridge between an L0-norm and a LASSO penalty, which empowers SHC-SSL without any a priori information. (D) Benchmarking of the sparse hierarchical clustering algorithms for feature selection precision and recall, as well as clustering performance, using 100 randomly generated simulation data sets. Silhouette measures consistency of the dissimilarity matrix based on discriminant features and ground-truth labels of clusters; concordance measures clustering performance after hierarchical clustering; and the adjusted Rand index (ARI) and purity measure clustering performance after Leiden clustering. For each metric, a higher value is better.