OKR-CELL Introduces Robust Single-Cell Foundation Model

Haoran Wang and colleagues on Jan. 9, 2026 preprint OKR-CELL, a cross-modal foundation model for single-cell multi-omics that was pretrained on 32 million cell-text pairs. The approach uses LLM-based retrieval-augmented generation to enrich cell descriptions and a Cross-modal Robust Alignment objective incorporating reliability scoring, curriculum learning, and coupled momentum contrastive learning. OKR-CELL achieves leading results across six tasks including clustering, annotation, batch correction, and zero-shot retrieval.