scLinguist: A pre-trained hyena-based foundation model for cross-modality translation in single-cell multi-omics

Single-cell multi-omics provides complementary insights into cellular states, heterogeneity, and regulatory programmes. However, paired assays remain costly, low-throughput, and technically challenging, whereas large-scale single-modality data such as scRNA-seq are abundant but do not capture protein-level biology. Here we present scLinguist, a foundation model for cross-modality translation introduces a three-stage framework: self-supervised pretraining on large-scale unimodal datasets to learn modality-specific representations, post-pretraining on limited paired data to capture cross-modality relationships, and inference to predict missing modalities (e.g., protein from RNA) in fine-tuning or zero-shot settings. Systematic benchmarking shows that scLinguist consistently outperforms state-of-the-art methods in predicting protein abundance from RNA across diverse biological contexts. It achieves high predictive performance while preserving cellular heterogeneity and further enables mechanistic and generalizable inference under simulated gene perturbations. Furthermore, scLinguist exhibits strong transferability across health states and datasets. By leveraging abundant unimodal data and minimizing dependence on paired assays, scLinguist establishes a scalable and versatile framework for cross-modality translation in single-cell analysis.