You Only Train Once: Differentiable Subset Selection for Omics Data
Researchers have developed YOTO, a novel end-to-end framework for identifying informative gene subsets from single-cell transcriptomic data. This approach integrates discrete gene selection and prediction within a single differentiable architecture, allowing the prediction task to directly guide gene selection. YOTO enforces sparsity, ensuring only selected genes contribute to inference and eliminating the need for separate downstream classifiers. The framework's multi-task learning design enables shared representations across objectives, improving generalization and performance on partially labeled datasets. AI
IMPACT Streamlines biomarker discovery and improves interpretability in omics data analysis.