MUSS is a multi‑scale, sequence‑based model for single‑cell gene regulation that preserves both DNA‑level sequence information and cell‑level regulatory states. The architecture leverages a frozen pre‑trained sequence encoder (e.g., a Transformer language model) to extract vector representations for chromatin peaks and genes directly from genomic DNA. These representations are then integrated via peak‑peak self‑attention to capture co‑accessibility among peaks and through distance‑biased peak‑gene interactions that encode the spatial relationship between a peak and its target gene.\n\nDuring pre‑training, MUSS employs peak masked language modeling, randomly masking portions of the peak sequences and predicting the original nucleotides, thereby learning a generic DNA grammar. The model is subsequently fine‑tuned on paired scATAC‑seq and scRNA‑seq data, aligning the learned embeddings with cell‑specific chromatin accessibility and gene expression profiles.\n\nMUSS achieves strong results on several downstream tasks:\n- Gene expression prediction: directly infers expression levels from peak embeddings for each cell.\n- Enhancer‑gene linking: ranks enhancer‑gene pairs using distance‑weighted interaction scores.\n- Motif‑guided regulatory prioritization: combines known transcription‑factor motifs with learned embeddings to prioritize regulatory elements.\n- Zero‑shot cell clustering: forms meaningful clusters even for unseen cell types.\n\nThe model also generalizes to genes and peaks that were absent during training, demonstrating robust zero‑shot prediction capabilities. All code, datasets, and model checkpoints are publicly available on GitHub.\n\nReview: MUSS successfully bridges sequence‑level language modeling with single‑cell multi‑omics, delivering high‑precision decoding of cell‑type‑specific regulatory programs and offering a scalable foundation for future gene‑regulation studies.