scLinguist

Single-cell multi-omics experiments can jointly profile RNA and surface protein in the same cell, but paired measurements remain scarce and expensive compared with the abundance of unimodal transcriptomic data. Predicting one modality from another, for example inferring protein abundance from a cell's transcriptome, would let researchers extend rich multi-omic analyses to the vast body of RNA-only datasets.

scLinguist, developed by researchers at Central South University together with the OmicsML group and released as a preprint in October 2025, is a foundation model built for this cross-modality translation problem. It frames the task as learning a shared "language" across omics layers, using a Hyena long-convolution backbone to capture dependencies across large feature sets and a staged training curriculum that separates learning within a modality from learning to map between modalities.

#Key Features

• Cross-modality translation: Predicts missing modalities, such as protein abundance from single-cell RNA, extending multi-omic analysis to unimodal data.
• Hyena backbone: Uses a Hyena long-convolution architecture in an encoder-decoder configuration to model long-range dependencies across features efficiently.
• Three-stage training: Combines self-supervised pretraining on large unpaired datasets, post-pretraining on limited paired data, and an inference stage that imputes the missing modality.
• Preserved heterogeneity and perturbation inference: Retains cellular heterogeneity in its predictions and supports mechanistic inference under gene perturbations, with transfer across datasets and biological contexts.
• Open release: Code is released under the MIT license alongside pretrained checkpoints.

#Technical Details

scLinguist adopts an encoder-decoder design with a Hyena long-convolution backbone. Its three-stage framework begins with modality-specific self-supervised pretraining on large unpaired corpora, reported as more than 15 million human cells for the RNA model and more than 11 million for the protein model, followed by post-pretraining on roughly 3 million paired RNA-protein cells to learn the cross-modality mapping, and finally an inference stage that predicts the absent modality for new cells. The authors release three checkpoints, RNA-pretrained, protein-pretrained, and the post-pretrained RNA-protein translator, under an MIT license. Benchmarking reports improved translation accuracy over prior methods while preserving cellular heterogeneity.

#Applications

The model lets researchers impute surface-protein profiles for RNA-only single-cell datasets, integrate and harmonize multi-omic measurements, and probe how predicted protein abundances shift under gene perturbations. Its reported transferability across datasets and tissues makes it useful for reanalyzing existing atlases and for extending CITE-seq-style analyses to experiments that measured only transcriptomes.

#Impact

scLinguist brings long-convolution Hyena architectures, previously explored mainly for genomic sequence modeling, to the problem of single-cell cross-modality translation, and pairs the approach with an open MIT-licensed release of code and checkpoints that supports independent use. As a preprint, its benchmark results await peer review.