CellPatch: a flexible and efficient framework for single-cell foundation models empowered by heuristic gene patching

Foundation models have advanced single-cell omics analysis but often remain too complex for practical use. We present CellPatch, a lightweight foundation model combining cross-attention with heuristic gene patching to reduce computational complexity during pretraining and fine-tuning while enabling inspection of gene-attention patterns. Across single-cell RNA-sequencing datasets spanning multiple organs and tissue states, CellPatch achieves state-of-the-art performance on diverse prediction tasks. Its modular design supports linear, adapter-based and task-specific architectures, enabling CellPatch to be adapted to diverse biological tasks.