Single-cell foundation models (scFMs) have increasingly relied on large‑scale transcriptomic pretraining, yet merely enlarging the dataset often yields diminishing returns while dramatically raising computational cost. Our scaling analysis revealed that incorporating biological knowledge—specifically cell‑level text annotations and gene‑level regulatory information—offers a more effective scaling dimension than data volume alone. Motivated by this insight, we introduce scKITE, a simple yet powerful scFM. scKITE integrates cell‑annotation and gene‑regulatory supervision into a shared transcriptomic Transformer encoder via lightweight auxiliary decoders that are active only during pretraining; after pretraining, the decoders are discarded, leaving a knowledge‑enriched encoder for downstream use. With just 179,067 pretraining samples (less than 0.5% of those used by prior strong scFMs), scKITE outperforms those models across diverse downstream tasks, highlighting knowledge‑enhanced pretraining as a promising paradigm for biologically grounded scFMs.
Review