Learning Minimal Gene Programs for Disease-Aligned Representations

Identifying small, interpretable gene sets that robustly capture disease-associated variation in single-cell transcriptomic data remains a central challenge for biological interpretation and experimental follow-up. In practice, commonly used differential expression and sparsity-based approaches often produce large, unstable gene lists that fail to generalize across patients due to strong donor-specific confounding. We study sparse gene selection for reconstructing donor-robust, disease-aligned cellular trajectories in real single-cell RNA-seq datasets. We introduce Sparse Linear Manifold Control (SLMC), a practical workflow that defines a disease-aligned score after removing donor-associated variation and selects minimal gene programs whose expression reconstructs this score. We focus on diagnosing the structure of the resulting reconstruction objective and evaluating selection strategies under realistic health data conditions. Across five human single-cell datasets spanning oncology and neurodegeneration, we find that the reconstruction objective exhibits strong diminishing returns, explaining why simple greedy selection methods perform well in practice. Under strict donor-held-out evaluation, greedy methods consistently outperform LASSO at small gene budgets and achieve accurate reconstruction with as few as 25 genes. Together, these results highlight how careful objective design and empirical evaluation enable robust and interpretable gene selection for disease-aligned representation learning in single-cell health data.