Identifying Donor-Robust Perturbation Targets via Sparse Manifold Control

Identifying small, donor-robust gene sets that capture disease-relevant variation in single-cell transcriptomics is a key bottleneck for mechanistic analysis and downstream perturbation modeling. We propose Sparse Linear Manifold Control (SLMC), which performs sparse gene selection for disease-aligned manifold reconstruction using donor-subspace orthogonalization in a structured PCA space. Across five diverse single-cell RNA-seq datasets, the resulting ridge reconstruction objective exhibits approximate diminishing returns, with rare and low-magnitude submodularity violations. We compare forward greedy, forward–backward greedy (FBG), beam search, and LASSO under fixed gene budgets in renal cell carcinoma and Alzheimer’s disease datasets. Greedy methods consistently outperform LASSO, recovering high-fidelity disease-aligned targets with as few as 25 genes, while FBG matches or exceeds beam search. Overall, objective structure, rather than algorithmic complexity, enables efficient and interpretable target prioritization in single-cell data.