Back to AI Research

AI Research

MUSS links DNA sequence and single-cell measurements to predict gene expression

Key Takeaways

  • MUSS combines DNA sequence and single-cell accessibility to predict gene expression, including for held-out sequences in four paired research datasets.
  • MUSS uses DNA-derived representations of genes and accessible regions.
  • The authors test expression prediction when particular genes, regions or cell types are held out.
  • The authors of the MUSS paper train a model to connect DNA sequence to the activity of genes in individual cell types.
  • They ask whether the same sequence-based representations can be used for genes and accessible DNA regions that were held out during task training.

The authors of the MUSS paper train a model to connect DNA sequence to the activity of genes in individual cell types. They ask whether the same sequence-based representations can be used for genes and accessible DNA regions that were held out during task training. The preprint was submitted to arXiv on November 5, 2026.

How MUSS represents a cell

The model derives representations for genes and accessible regions, called peaks, from a frozen DNA sequence encoder. It uses self-attention among peaks and distance-biased interactions between peaks and genes. The first training stage asks the model to recover masked peaks. The next stage uses paired single-cell chromatin accessibility and gene expression measurements to connect that context to expression.
Because each region is represented from its sequence, the model can encode an unseen region without adding a previously learned identifier for it. The learned interactions can then use the new representation alongside the other peaks and genes in the cell.

What the authors measure on held-out data

The expression experiments cross sequence splits with cell-type splits across four paired datasets. The authors report the best or tied-best per-cell-type aggregate Pearson correlation for six of eight unseen-sequence settings. Another model, scooby, scores higher on the concordance metric for the Bone dataset in the unseen-sequence condition. The differences between metrics matter when using the results to choose a model for a particular dataset.

What the other tasks show

The authors also evaluate enhancer-to-gene links and zero-shot cell clustering. In the held-out PBMC169K batch, they report an adjusted Rand index of 0.5377 for MUSS and 0.4835 for ChromFound. That result concerns grouping cells in the benchmark. Readers interested in regulatory mechanisms should distinguish such predictive scores from direct experimental evidence about what happens when a regulatory sequence is changed.

Comments