#18 · ML / AI methods

Recovering cell-type expression from sparse, mixed, and noisy transcriptomic measurements

The dominant thread in this set is learned representations that align, embed, or integrate transcriptomic and multi-omic measurements so that structure can be recovered from heterogeneous or sparse inputs. STORM is the closest match to the recovery theme: it jointly performs batch alignment, cell type deconvolution, and gene imputation on spatial transcriptomic slices with varying geometry, using an interpretable irregular tensor factorization. 3d-OT and SMART extend spatial multi-omic integration, with 3d-OT addressing nonrigid alignment of heterogeneous slices through soft optimal transport and SMART using graph neural networks with metric learning to identify anatomical regions. HSSE applies persistent sheaf Laplacians to derive multiscale features for scRNA-seq. The remaining items, Dynomap, the Boosting Autoencoder for ligand-receptor interaction data, and Triad-LMF for cancer subtype classification, address adjacent problems rather than imputation or deconvolution directly.