#30 · Single-cell

Constrained optimization for marker selection, deconvolution and cell identity

The dominant thread is explicit formulation of inference as an optimization problem over single-cell, spatial and bulk accessibility data. WAD deconvolves cell-type proportions from bulk ATAC-seq via wavelet-based L1-minimal linear programming, while SepSolve selects small marker sets through a linear program that enforces c-separation of cell types. geneCover offers a label-free route to minimally redundant marker panels from gene-gene correlations, and SIID uses joint nonnegative factorization to impute and deconvolve genes across paired Xenium and Visium platforms. ALPINE applies supervised nonnegative matrix factorization to separate technical from phenotypic factors in multicondition data. The subspace clustering method models each cell type as an affine subspace of linear constraints on scRNA-seq profiles. What is new is the constraint or objective itself, such as L1 reconstruction loss, c-separation or affine subspace membership, used as the basis of identity and composition inference.